# Questions on Logistic Regression using R

AllenFlower

DATA MINING FOR BUSINESS ANALYTICS

DATA MINING FOR BUSINESS ANALYTICS

Concepts, Techniques, and Applications in R

Galit Shmueli

Peter C. Bruce

Inbal Yahav

Nitin R. Patel

Kenneth C. Lichtendahl, Jr.

This edition first published 2018

© 2018 John Wiley & Sons, Inc.

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by law. Advice on how to obtain permission to reuse material from this title is available at http://www.wiley.com/go/permissions.

The right of Galit Shmueli, Peter C. Bruce, Inbal Yahav, Nitin R. Patel, and Kenneth C. Lichtendahl Jr. to be identified as the authors of this work has been asserted in accordance with law.

Registered Offices John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA

Editorial Office 111 River Street, Hoboken, NJ 07030, USA

For details of our global editorial offices, customer services, and more information about Wiley products visit us at www.wiley.com.

Wiley also publishes its books in a variety of electronic formats and by print-on-demand. Some content that appears in standard print versions of this book may not be available in other formats.

Limit of Liability/Disclaimer of Warranty The publisher and the authors make no representations or warranties with respect to the accuracy or completeness of the contents of this work and specifically disclaim all warranties; including without limitation any implied warranties of fitness for a particular purpose. This work is sold with the understanding that the publisher is not engaged in rendering professional services. The advice and strategies contained herein may not be suitable for every situation. In view of on-going research, equipment modifications, changes in governmental regulations, and the constant flow of information relating to the use of experimental reagents, equipment, and devices, the reader is urged to review and evaluate the information provided in the package insert or instructions for each chemical, piece of equipment, reagent, or device for, among other things, any changes in the instructions or indication of usage and for added warnings and precautions. The fact that an organization or website is referred to in this work as a citation and/or potential source of further information does not mean that the author or the publisher endorses the information the organization or website may provide or recommendations it may make. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this works was written and when it is read. No warranty may be created or extended by any promotional statements for this work. Neither the publisher nor the author shall be liable for any damages arising here from.

Library of Congress Cataloging-in-Publication Data applied for

Hardback: 9781118879368

Cover Design: Wiley Cover Image: © Achim Mittler, Frankfurt am Main/Gettyimages

Set in 11.5/14.5pt BemboStd by Aptara Inc., New Delhi, India Printed in the United States of America.

10 9 8 7 6 5 4 3 2 1

The beginning of wisdom is this:

Get wisdom, and whatever else you get, get insight.

– Proverbs 4:7

Contents

Foreword by Gareth James xix

Foreword by Ravi Bapna xxi

Preface to the R Edition xxiii

Acknowledgments xxvii

PART I PRELIMINARIES CHAPTER 1 Introduction 3

1.1 What Is Business Analytics? . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 What Is Data Mining? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.3 Data Mining and Related Terms . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.4 Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.5 Data Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.6 Why Are There So Many Different Methods? . . . . . . . . . . . . . . . . . . . 8 1.7 Terminology and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.8 Road Maps to This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Order of Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

CHAPTER 2 Overview of the Data Mining Process 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2 Core Ideas in Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Association Rules and Recommendation Systems . . . . . . . . . . . . . . . . . 16 Predictive Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Data Reduction and Dimension Reduction . . . . . . . . . . . . . . . . . . . . 17 Data Exploration and Visualization . . . . . . . . . . . . . . . . . . . . . . . . 17 Supervised and Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . 18

2.3 The Steps in Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.4 Preliminary Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

Organization of Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Predicting Home Values in the West Roxbury Neighborhood . . . . . . . . . . . 21

vii

viii CONTENTS

Loading and Looking at the Data in R . . . . . . . . . . . . . . . . . . . . . . 22 Sampling from a Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Oversampling Rare Events in Classification Tasks . . . . . . . . . . . . . . . . . 25 Preprocessing and Cleaning the Data . . . . . . . . . . . . . . . . . . . . . . . 26

2.5 Predictive Power and Overfitting . . . . . . . . . . . . . . . . . . . . . . . . . 33 Overfitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Creation and Use of Data Partitions . . . . . . . . . . . . . . . . . . . . . . . 35

2.6 Building a Predictive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Modeling Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.7 Using R for Data Mining on a Local Machine . . . . . . . . . . . . . . . . . . . 43 2.8 Automating Data Mining Solutions . . . . . . . . . . . . . . . . . . . . . . . . 43

Data Mining Software: The State of the Market (by Herb Edelstein) . . . . . . . . 45 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

PART II DATA EXPLORATION AND DIMENSION REDUCTION CHAPTER 3 Data Visualization 55

3.1 Uses of Data Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 Base R or ggplot? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.2 Data Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 Example 1: Boston Housing Data . . . . . . . . . . . . . . . . . . . . . . . . 57 Example 2: Ridership on Amtrak Trains . . . . . . . . . . . . . . . . . . . . . . 59

3.3 Basic Charts: Bar Charts, Line Graphs, and Scatter Plots . . . . . . . . . . . . . 59 Distribution Plots: Boxplots and Histograms . . . . . . . . . . . . . . . . . . . 61 Heatmaps: Visualizing Correlations and Missing Values . . . . . . . . . . . . . . 64

3.4 Multidimensional Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . 67 Adding Variables: Color, Size, Shape, Multiple Panels, and Animation . . . . . . . 67 Manipulations: Rescaling, Aggregation and Hierarchies, Zooming, Filtering . . . . 70 Reference: Trend Lines and Labels . . . . . . . . . . . . . . . . . . . . . . . . 74 Scaling up to Large Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Multivariate Plot: Parallel Coordinates Plot . . . . . . . . . . . . . . . . . . . . 75 Interactive Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.5 Specialized Visualizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Visualizing Networked Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Visualizing Hierarchical Data: Treemaps . . . . . . . . . . . . . . . . . . . . . 82 Visualizing Geographical Data: Map Charts . . . . . . . . . . . . . . . . . . . . 83

3.6 Summary: Major Visualizations and Operations, by Data Mining Goal . . . . . . . 86 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Time Series Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

CHAPTER 4 Dimension Reduction 91

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 4.2 Curse of Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

CONTENTS ix

4.3 Practical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Example 1: House Prices in Boston . . . . . . . . . . . . . . . . . . . . . . . 93

4.4 Data Summaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

Aggregation and Pivot Tables . . . . . . . . . . . . . . . . . . . . . . . . . . 96

4.5 Correlation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.6 Reducing the Number of Categories in Categorical Variables . . . . . . . . . . . 99

4.7 Converting a Categorical Variable to a Numerical Variable . . . . . . . . . . . . 99

4.8 Principal Components Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 101

Example 2: Breakfast Cereals . . . . . . . . . . . . . . . . . . . . . . . . . . 101

Principal Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

Normalizing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Using Principal Components for Classification and Prediction . . . . . . . . . . . 109

4.9 Dimension Reduction Using Regression Models . . . . . . . . . . . . . . . . . . 111

4.10 Dimension Reduction Using Classification and Regression Trees . . . . . . . . . . 111

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

PART III PERFORMANCE EVALUATION

CHAPTER 5 Evaluating Predictive Performance 117

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

5.2 Evaluating Predictive Performance . . . . . . . . . . . . . . . . . . . . . . . . 118

Naive Benchmark: The Average . . . . . . . . . . . . . . . . . . . . . . . . . 118

Prediction Accuracy Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 119

Comparing Training and Validation Performance . . . . . . . . . . . . . . . . . 121

Lift Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

5.3 Judging Classifier Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Benchmark: The Naive Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Class Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

The Confusion (Classification) Matrix . . . . . . . . . . . . . . . . . . . . . . . 124

Using the Validation Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Accuracy Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Propensities and Cutoff for Classification . . . . . . . . . . . . . . . . . . . . . 127

Performance in Case of Unequal Importance of Classes . . . . . . . . . . . . . . 131

Asymmetric Misclassification Costs . . . . . . . . . . . . . . . . . . . . . . . . 133

Generalization to More Than Two Classes . . . . . . . . . . . . . . . . . . . . . 135

5.4 Judging Ranking Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 136

Lift Charts for Binary Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

Decile Lift Charts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

Beyond Two Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Lift Charts Incorporating Costs and Benefits . . . . . . . . . . . . . . . . . . . 139

Lift as a Function of Cutoff . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5.5 Oversampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

Oversampling the Training Set . . . . . . . . . . . . . . . . . . . . . . . . . . 144

x CONTENTS

Evaluating Model Performance Using a Non-oversampled Validation Set . . . . . . 144 Evaluating Model Performance if Only Oversampled Validation Set Exists . . . . . 144

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

PART IV PREDICTION AND CLASSIFICATION METHODS CHAPTER 6 Multiple Linear Regression 153

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 6.2 Explanatory vs. Predictive Modeling . . . . . . . . . . . . . . . . . . . . . . . 154 6.3 Estimating the Regression Equation and Prediction . . . . . . . . . . . . . . . . 156

Example: Predicting the Price of Used Toyota Corolla Cars . . . . . . . . . . . . 156 6.4 Variable Selection in Linear Regression . . . . . . . . . . . . . . . . . . . . . 161

Reducing the Number of Predictors . . . . . . . . . . . . . . . . . . . . . . . 161 How to Reduce the Number of Predictors . . . . . . . . . . . . . . . . . . . . . 162

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

CHAPTER 7 k-Nearest Neighbors (kNN) 173

7.1 The k-NN Classifier (Categorical Outcome) . . . . . . . . . . . . . . . . . . . . 173 Determining Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 Classification Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 Example: Riding Mowers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 Choosing k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Setting the Cutoff Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 k-NN with More Than Two Classes . . . . . . . . . . . . . . . . . . . . . . . . 180 Converting Categorical Variables to Binary Dummies . . . . . . . . . . . . . . . 180

7.2 k-NN for a Numerical Outcome . . . . . . . . . . . . . . . . . . . . . . . . . . 180 7.3 Advantages and Shortcomings of k-NN Algorithms . . . . . . . . . . . . . . . . 182 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

CHAPTER 8 The Naive Bayes Classifier 187

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Cutoff Probability Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Conditional Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Example 1: Predicting Fraudulent Financial Reporting . . . . . . . . . . . . . . 188

8.2 Applying the Full (Exact) Bayesian Classifier . . . . . . . . . . . . . . . . . . . 189 Using the “Assign to the Most Probable Class” Method . . . . . . . . . . . . . . 190 Using the Cutoff Probability Method . . . . . . . . . . . . . . . . . . . . . . . 190 Practical Difficulty with the Complete (Exact) Bayes Procedure . . . . . . . . . . 190 Solution: Naive Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 The Naive Bayes Assumption of Conditional Independence . . . . . . . . . . . . 192 Using the Cutoff Probability Method . . . . . . . . . . . . . . . . . . . . . . . 192 Example 2: Predicting Fraudulent Financial Reports, Two Predictors . . . . . . . 193 Example 3: Predicting Delayed Flights . . . . . . . . . . . . . . . . . . . . . . 194

8.3 Advantages and Shortcomings of the Naive Bayes Classifier . . . . . . . . . . . 199 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

CONTENTS xi

CHAPTER 9 Classification and Regression Trees 205

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

9.2 Classification Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

Recursive Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

Example 1: Riding Mowers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

Measures of Impurity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

Tree Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

Classifying a New Record . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

9.3 Evaluating the Performance of a Classification Tree . . . . . . . . . . . . . . . . 215

Example 2: Acceptance of Personal Loan . . . . . . . . . . . . . . . . . . . . . 215

9.4 Avoiding Overfitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216

Stopping Tree Growth: Conditional Inference Trees . . . . . . . . . . . . . . . . 221

Pruning the Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222

Cross-Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222

Best-Pruned Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

9.5 Classification Rules from Trees . . . . . . . . . . . . . . . . . . . . . . . . . . 226

9.6 Classification Trees for More Than Two Classes . . . . . . . . . . . . . . . . . . 227

9.7 Regression Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

Measuring Impurity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

Evaluating Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

9.8 Improving Prediction: Random Forests and Boosted Trees . . . . . . . . . . . . 229

Random Forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

Boosted Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

9.9 Advantages and Weaknesses of a Tree . . . . . . . . . . . . . . . . . . . . . . 232

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

CHAPTER 10 Logistic Regression 237

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

10.2 The Logistic Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . 239

10.3 Example: Acceptance of Personal Loan . . . . . . . . . . . . . . . . . . . . . . 240

Model with a Single Predictor . . . . . . . . . . . . . . . . . . . . . . . . . . 241

Estimating the Logistic Model from Data: Computing Parameter Estimates . . . . 243

Interpreting Results in Terms of Odds (for a Profiling Goal) . . . . . . . . . . . . 244

10.4 Evaluating Classification Performance . . . . . . . . . . . . . . . . . . . . . . 247

Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

10.5 Example of Complete Analysis: Predicting Delayed Flights . . . . . . . . . . . . 250

Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

Model-Fitting and Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Model Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Model Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

10.6 Appendix: Logistic Regression for Profiling . . . . . . . . . . . . . . . . . . . . 259

Appendix A: Why Linear Regression Is Problematic for a Categorical Outcome . . . 259

xii CONTENTS

Appendix B: Evaluating Explanatory Power . . . . . . . . . . . . . . . . . . . . 261 Appendix C: Logistic Regression for More Than Two Classes . . . . . . . . . . . . 264

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

CHAPTER 11 Neural Nets 271

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 11.2 Concept and Structure of a Neural Network . . . . . . . . . . . . . . . . . . . . 272 11.3 Fitting a Network to Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273

Example 1: Tiny Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 Computing Output of Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 Preprocessing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 Training the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 Example 2: Classifying Accident Severity . . . . . . . . . . . . . . . . . . . . . 282 Avoiding Overfitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Using the Output for Prediction and Classification . . . . . . . . . . . . . . . . 283

11.4 Required User Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 11.5 Exploring the Relationship Between Predictors and Outcome . . . . . . . . . . . 287 11.6 Advantages and Weaknesses of Neural Networks . . . . . . . . . . . . . . . . . 288 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290

CHAPTER 12 Discriminant Analysis 293

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Example 1: Riding Mowers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 Example 2: Personal Loan Acceptance . . . . . . . . . . . . . . . . . . . . . . 294

12.2 Distance of a Record from a Class . . . . . . . . . . . . . . . . . . . . . . . . 296 12.3 Fisher’s Linear Classification Functions . . . . . . . . . . . . . . . . . . . . . . 297 12.4 Classification Performance of Discriminant Analysis . . . . . . . . . . . . . . . 300 12.5 Prior Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 12.6 Unequal Misclassification Costs . . . . . . . . . . . . . . . . . . . . . . . . . 302 12.7 Classifying More Than Two Classes . . . . . . . . . . . . . . . . . . . . . . . . 303

Example 3: Medical Dispatch to Accident Scenes . . . . . . . . . . . . . . . . . 303 12.8 Advantages and Weaknesses . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

CHAPTER 13 Combining Methods: Ensembles and Uplift Modeling 311

13.1 Ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 Why Ensembles Can Improve Predictive Power . . . . . . . . . . . . . . . . . . 312 Simple Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314 Bagging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 Bagging and Boosting in R . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 Advantages and Weaknesses of Ensembles . . . . . . . . . . . . . . . . . . . . 315

13.2 Uplift (Persuasion) Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 317 A-B Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318

CONTENTS xiii

Uplift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318 Gathering the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319 A Simple Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320 Modeling Individual Uplift . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 Computing Uplift with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 Using the Results of an Uplift Model . . . . . . . . . . . . . . . . . . . . . . . 322

13.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325

PART V MINING RELATIONSHIPS AMONG RECORDS CHAPTER 14 Association Rules and Collaborative Filtering 329

14.1 Association Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 Discovering Association Rules in Transaction Databases . . . . . . . . . . . . . 330 Example 1: Synthetic Data on Purchases of Phone Faceplates . . . . . . . . . . 330 Generating Candidate Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 The Apriori Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 Selecting Strong Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 Data Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 The Process of Rule Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 336 Interpreting the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 Rules and Chance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 Example 2: Rules for Similar Book Purchases . . . . . . . . . . . . . . . . . . . 340

14.2 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342 Data Type and Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 Example 3: Netflix Prize Contest . . . . . . . . . . . . . . . . . . . . . . . . . 343 User-Based Collaborative Filtering: “People Like You” . . . . . . . . . . . . . . 344 Item-Based Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . 347 Advantages and Weaknesses of Collaborative Filtering . . . . . . . . . . . . . . 348 Collaborative Filtering vs. Association Rules . . . . . . . . . . . . . . . . . . . 349

14.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352

CHAPTER 15 Cluster Analysis 357

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357 Example: Public Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

15.2 Measuring Distance Between Two Records . . . . . . . . . . . . . . . . . . . . 361 Euclidean Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 Normalizing Numerical Measurements . . . . . . . . . . . . . . . . . . . . . . 362 Other Distance Measures for Numerical Data . . . . . . . . . . . . . . . . . . . 362 Distance Measures for Categorical Data . . . . . . . . . . . . . . . . . . . . . . 365 Distance Measures for Mixed Data . . . . . . . . . . . . . . . . . . . . . . . . 366

15.3 Measuring Distance Between Two Clusters . . . . . . . . . . . . . . . . . . . . 366 Minimum Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 Maximum Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366

xiv CONTENTS

Average Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

Centroid Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

15.4 Hierarchical (Agglomerative) Clustering . . . . . . . . . . . . . . . . . . . . . 368

Single Linkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

Complete Linkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

Average Linkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

Centroid Linkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

Ward’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

Dendrograms: Displaying Clustering Process and Results . . . . . . . . . . . . . 371

Validating Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373

Limitations of Hierarchical Clustering . . . . . . . . . . . . . . . . . . . . . . 375

15.5 Non-Hierarchical Clustering: The k-Means Algorithm . . . . . . . . . . . . . . . 376

Choosing the Number of Clusters (k) . . . . . . . . . . . . . . . . . . . . . . . 377

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382

PART VI FORECASTING TIME SERIES

CHAPTER 16 Handling Time Series 387

16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387

16.2 Descriptive vs. Predictive Modeling . . . . . . . . . . . . . . . . . . . . . . . 389

16.3 Popular Forecasting Methods in Business . . . . . . . . . . . . . . . . . . . . . 389

Combining Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389

16.4 Time Series Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390

Example: Ridership on Amtrak Trains . . . . . . . . . . . . . . . . . . . . . . . 390

16.5 Data-Partitioning and Performance Evaluation . . . . . . . . . . . . . . . . . . 395

Benchmark Performance: Naive Forecasts . . . . . . . . . . . . . . . . . . . . 395

Generating Future Forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . 396

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398

CHAPTER 17 Regression-Based Forecasting 401

17.1 A Model with Trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401

Linear Trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401

Exponential Trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405

Polynomial Trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407

17.2 A Model with Seasonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407

17.3 A Model with Trend and Seasonality . . . . . . . . . . . . . . . . . . . . . . . 411

17.4 Autocorrelation and ARIMA Models . . . . . . . . . . . . . . . . . . . . . . . . 412

Computing Autocorrelation . . . . . . . . . . . . . . . . . . . . . . . . . . . 413

Improving Forecasts by Integrating Autocorrelation Information . . . . . . . . . 416

Evaluating Predictability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422

CONTENTS xv

CHAPTER 18 Smoothing Methods 433

18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433 18.2 Moving Average . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434

Centered Moving Average for Visualization . . . . . . . . . . . . . . . . . . . . 434 Trailing Moving Average for Forecasting . . . . . . . . . . . . . . . . . . . . . 435 Choosing Window Width (w) . . . . . . . . . . . . . . . . . . . . . . . . . . . 439

18.3 Simple Exponential Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . 439 Choosing Smoothing Parameter α . . . . . . . . . . . . . . . . . . . . . . . . 440 Relation Between Moving Average and Simple Exponential Smoothing . . . . . . 440

18.4 Advanced Exponential Smoothing . . . . . . . . . . . . . . . . . . . . . . . . 442 Series with a Trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442 Series with a Trend and Seasonality . . . . . . . . . . . . . . . . . . . . . . . 443 Series with Seasonality (No Trend) . . . . . . . . . . . . . . . . . . . . . . . . 443

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446

PART VII DATA ANALYTICS CHAPTER 19 Social Network Analytics 455

19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455 19.2 Directed vs. Undirected Networks . . . . . . . . . . . . . . . . . . . . . . . . 457 19.3 Visualizing and Analyzing Networks . . . . . . . . . . . . . . . . . . . . . . . 458

Graph Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458 Edge List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460 Adjacency Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461 Using Network Data in Classification and Prediction . . . . . . . . . . . . . . . 461

19.4 Social Data Metrics and Taxonomy . . . . . . . . . . . . . . . . . . . . . . . . 462 Node-Level Centrality Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 463 Egocentric Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463 Network Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465

19.5 Using Network Metrics in Prediction and Classification . . . . . . . . . . . . . . 467 Link Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467 Entity Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468

19.6 Collecting Social Network Data with R . . . . . . . . . . . . . . . . . . . . . . 471 19.7 Advantages and Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . 474 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476

CHAPTER 20 Text Mining 479

20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479 20.2 The Tabular Representation of Text: Term-Document Matrix and “Bag-of-Words” . 480 20.3 Bag-of-Words vs. Meaning Extraction at Document Level . . . . . . . . . . . . . 481 20.4 Preprocessing the Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482

Tokenization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484 Text Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485

xvi CONTENTS

Presence/Absence vs. Frequency . . . . . . . . . . . . . . . . . . . . . . . . . 487

Term Frequency–Inverse Document Frequency (TF-IDF) . . . . . . . . . . . . . . 487

From Terms to Concepts: Latent Semantic Indexing . . . . . . . . . . . . . . . 488

Extracting Meaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489

20.5 Implementing Data Mining Methods . . . . . . . . . . . . . . . . . . . . . . . 489

20.6 Example: Online Discussions on Autos and Electronics . . . . . . . . . . . . . . 490

Importing and Labeling the Records . . . . . . . . . . . . . . . . . . . . . . . 490

Text Preprocessing in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491

Producing a Concept Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 491

Fitting a Predictive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492

Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492

20.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495

PART VIII CASES CHAPTER 21 Cases 499

21.1 Charles Book Club . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499

The Book Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499

Database Marketing at Charles . . . . . . . . . . . . . . . . . . . . . . . . . . 500

Data Mining Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502

Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504

21.2 German Credit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506

Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507

21.3 Tayko Software Cataloger . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510

The Mailing Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510

Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512

21.4 Political Persuasion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513

Predictive Analytics Arrives in US Politics . . . . . . . . . . . . . . . . . . . . 513

Political Targeting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514

Uplift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515

Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516

21.5 Taxi Cancellations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517

Business Situation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517

Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517

21.6 Segmenting Consumers of Bath Soap . . . . . . . . . . . . . . . . . . . . . . . 518

Business Situation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518

Key Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519

CONTENTS xvii

Measuring Brand Loyalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521

21.7 Direct-Mail Fundraising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523

21.8 Catalog Cross-Selling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524

21.9 Predicting Bankruptcy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 Predicting Corporate Bankruptcy . . . . . . . . . . . . . . . . . . . . . . . . . 525 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526

21.10 Time Series Case: Forecasting Public Transportation Demand . . . . . . . . . . . 528 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Available Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Assignment Goal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529 Tips and Suggested Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529

References 531

Data Files Used in the Book 533

Index 535

Foreword by Gareth James

T he field of statistics has existed in one form or another for 200 years, and bythe second half of the 20th century had evolved into a well-respected and essential academic discipline. However, its prominence expanded rapidly in the 1990s with the explosion of new, and enormous, data sources. For the first part of this century, much of this attention was focused on biological applications, in particular, genetics data generated as a result of the sequencing of the human genome. However, the last decade has seen a dramatic increase in the availability of data in the business disciplines, and a corresponding interest in business-related statistical applications.

The impact has been profound. Ten years ago, when I was able to attract a full class of MBA students to my new statistical learning elective, my colleagues were astonished because our department struggled to fill most electives. Today, we offer a Masters in Business Analytics, which is the largest specialized masters program in the school and has application volume rivaling those of our MBA programs. Our department’s faculty size and course offerings have increased dramatically, yet the MBA students are still complaining that the classes are all full. Google’s chief economist, Hal Varian, was indeed correct in 2009 when he stated that “the sexy job in the next 10 years will be statisticians.”

This demand is driven by a simple, but undeniable, fact. Business analyt- ics solutions have produced significant and measurable improvements in business performance, on multiple dimensions and in numerous settings, and as a result, there is a tremendous demand for individuals with the requisite skill set. How- ever, training students in these skills is challenging given that, in addition to the obvious required knowledge of statistical methods, they need to understand business-related issues, possess strong communication skills, and be comfortable dealing with multiple computational packages. Most statistics texts concentrate on abstract training in classical methods, without much emphasis on practical, let alone business, applications.

This book has by far the most comprehensive review of business analytics methods that I have ever seen, covering everything from classical approaches such as linear and logistic regression, through to modern methods like neural

xix

xx FOREWORD BY GARETH JAMES

networks, bagging and boosting, and even much more business specific proce- dures such as social network analysis and text mining. If not the bible, it is at the least a definitive manual on the subject. However, just as important as the list of topics, is the way that they are all presented in an applied fashion using business applications. Indeed the last chapter is entirely dedicated to 10 separate cases where business analytics approaches can be applied.

In this latest edition, the authors have added an important new dimension in the form of the R software package. Easily the most widely used and influ- ential open source statistical software, R has become the go-to tool for such purposes. With literally hundreds of freely available add-on packages, R can be used for almost any business analytics related problem. The book provides detailed descriptions and code involving applications of R in numerous business settings, ensuring that the reader will actually be able to apply their knowledge to real-life problems.

We recently introduced a business analytics course into our required MBA core curriculum and I intend to make heavy use of this book in developing the syllabus. I’m confident that it will be an indispensable tool for any such course.

GARETH JAMES

Marshall School of Business, University of Southern California, 2017

Foreword by Ravi Bapna

D ata is the new gold—and mining this gold to create business value in today’scontext of a highly networked and digital society requires a skillset that we haven’t traditionally delivered in business or statistics or engineering programs on their own. For those businesses and organizations that feel overwhelmed by today’s Big Data, the phrase you ain’t seen nothing yet comes to mind. Yester- day’s three major sources of Big Data—the 20+ years of investment in enterprise systems (ERP, CRM, SCM, …), the 3 billion plus people on the online social grid, and the close to 5 billion people carrying increasingly sophisticated mobile devices—are going to be dwarfed by tomorrow’s smarter physical ecosystems fueled by the Internet of Things (IoT) movement.

The idea that we can use sensors to connect physical objects such as homes, automobiles, roads, even garbage bins and streetlights, to digitally optimized systems of governance goes hand in glove with bigger data and the need for deeper analytical capabilities. We are not far away from a smart refrigerator sensing that you are short on, say, eggs, populating your grocery store’s mobile app’s shopping list, and arranging a Task Rabbit to do a grocery run for you. Or the refrigerator negotiating a deal with an Uber driver to deliver an evening meal to you. Nor are we far away from sensors embedded in roads and vehicles that can compute traffic congestion, track roadway wear and tear, record vehicle use and factor these into dynamic usage-based pricing, insurance rates, and even taxation. This brave new world is going to be fueled by analytics and the ability to harness data for competitive advantage.

Business Analytics is an emerging discipline that is going to help us ride this new wave. This new Business Analytics discipline requires individuals who are grounded in the fundamentals of business such that they know the right questions to ask, who have the ability to harness, store, and optimally process vast datasets from a variety of structured and unstructured sources, and who can then use an array of techniques from machine learning and statistics to uncover new insights for decision-making. Such individuals are a rare commodity today, but their creation has been the focus of this book for a decade now. This book’s forte is that it relies on explaining the core set of concepts required for today’s business analytics professionals using real-world data-rich cases in a hands-on manner,

xxi

xxii FOREWORD BY RAVI BAPNA

without sacrificing academic rigor. It provides a modern day foundation for Business Analytics, the notion of linking the x’s to the y’s of interest in a predictive sense. I say this with the confidence of someone who was probably the first adopter of the zeroth edition of this book (Spring 2006 at the Indian School of Business).

I can’t say enough about the long-awaited R edition. R is my go-to platform for analytics these days. It’s also used by a wide variety of instructors in our MS- Business Analytics program. The open-innovation paradigm used by R is one key part of the analytics perfect storm, the other components being the advances in computing and the business appetite for data-driven decision-making.

I look forward to using the book in multiple fora, in executive education, in MBA classrooms, in MS-Business Analytics programs, and in Data Science bootcamps. I trust you will too!

RAVI BAPNA

Carlson School of Management, University of Minnesota, 2017

Preface to the R Edition

T his textbook first appeared in early 2007 and has been used by numerousstudents and practitioners and in many courses, ranging from dedicated data mining classes to more general business analytics courses (including our own experience teaching this material both online and in person for more than 10 years). The first edition, based on the Excel add-in XLMiner, was followed by two more XLMiner editions, a JMP edition, and now this R edition, with its companion website, www.dataminingbook.com.

This new R edition, which relies on the free and open-source R soft- ware, presents output from R, as well as the code used to produce that output, including specification of a variety of packages and functions. Unlike computer- science or statistics-oriented textbooks, the focus in this book is on data mining concepts, and how to implement the associated algorithms in R. We assume a basic facility with R.

For this R edition, two new co-authors stepped on board—Inbal Yahav and Casey Lichtendahl—bringing both expertise teaching business analytics courses using R and data mining consulting experience in business and government. Such practical experience is important, since the open-source nature of R soft- ware makes available a plethora of approaches, packages, and functions available for data mining. Given the main goal of this book—to introduce data min- ing concepts using R software for illustration—our challenge was to choose an R code cocktail that supports highlighting the important concepts. In addi- tion to providing R code and output, this edition also incorporates updates and new material based on feedback from instructors teaching MBA, undergraduate, diploma, and executive courses, and from their students as well.

One update, compared to the first two editions of the book, is the title: we now use Business Analytics in place of Business Intelligence. This reflects the change in terminology since the second edition: Business Intelligence today refers mainly to reporting and data visualization (“what is happening now”), while Business Analytics has taken over the “advanced analytics,” which include predictive analytics and data mining. In this new edition, we therefore use the updated terms.

xxiii

xxiv PREFACE TO THE R EDITION

This R edition includes the material that was recently added in the third edition of the original (XLMiner-based) book:

• Social network analysis

• Text mining

• Ensembles

• Uplift modeling

• Collaborative filtering

Since the appearance of the (XLMiner-based) second edition, the landscape of the courses using the textbook has greatly expanded: whereas initially, the book was used mainly in semester-long elective MBA-level courses, it is now used in a variety of courses in Business Analytics degrees and certificate programs, ranging from undergraduate programs, to post-graduate and executive education programs. Courses in such programs also vary in their duration and coverage. In many cases, this textbook is used across multiple courses. The book is designed to continue supporting the general “Predictive Analytics” or “Data Mining” course as well as supporting a set of courses in dedicated business analytics programs.

A general “Business Analytics,” “Predictive Analytics,” or “Data Mining” course, common in MBA and undergraduate programs as a one-semester elec- tive, would cover Parts I–III, and choose a subset of methods from Parts IV and V. Instructors can choose to use cases as team assignments, class discussions, or projects. For a two-semester course, Part VI might be considered, and we recommend introducing the new Part VII (Data Analytics).

For a set of courses in a dedicated business analytics program, here are a few courses that have been using our book:

Predictive Analytics: Supervised Learning In a dedicated Business Analytics program, the topic of Predictive Analytics is typically instructed across a set of courses. The first course would cover Parts I–IV and instruc- tors typically choose a subset of methods from Part IV according to the course length. We recommend including the new Chapter 13 in such a course, as well as the new “Part VII: Data Analytics.”

Predictive Analytics: Unsupervised Learning This course introduces data exploration and visualization, dimension reduction, mining relation- ships, and clustering (Parts III and V). If this course follows the Predictive Analytics: Supervised Learning course, then it is useful to examine examples and approaches that integrate unsupervised and supervised learning, such as the new part on “Data Analytics.”

Forecasting Analytics A dedicated course on time series forecasting would rely on Part VI.

PREFACE TO THE R EDITION xxv

Advanced Analytics A course that integrates the learnings from Predictive Analytics (supervised and unsupervised learning). Such a course can focus on Part VII: Data Analytics, where social network analytics and text mining are introduced. Some instructors choose to use the Cases (Chapter 21) in such a course.

In all courses, we strongly recommend including a project component, where data are either collected by students according to their interest or pro- vided by the instructor (e.g., from the many data mining competition datasets available). From our experience and other instructors’ experience, such projects enhance the learning and provide students with an excellent opportunity to understand the strengths of data mining and the challenges that arise in the process.

Acknowledgments

We thank the many people who assisted us in improving the first three edi- tions of the initial XLMiner version of this book and the JMP edition, as well as those who helped with comments on early drafts of this R edition. Anthony Babinec, who has been using earlier editions of this book for years in his data mining courses at Statistics.com, provided us with detailed and expert correc- tions. Dan Toy and John Elder IV greeted our project with early enthusiasm and provided detailed and useful comments on initial drafts. Ravi Bapna, who used an early draft in a data mining course at the Indian School of Business, has provided invaluable comments and helpful suggestions since the book’s start.

Many of the instructors, teaching assistants, and students using earlier edi- tions of the book have contributed invaluable feedback both directly and indi- rectly, through fruitful discussions, learning journeys, and interesting data min- ing projects that have helped shape and improve the book. These include MBA students from the University of Maryland, MIT, the Indian School of Business, National Tsing Hua University, and Statistics.com. Instructors from many uni- versities and teaching programs, too numerous to list, have supported and helped improve the book since its inception.

Several professors have been especially helpful with this R edition: Hayri Tongarlak, Prashant Joshi, Jay Annadatha, Roger Bohn, and Sridhar Vaithianathan provided detailed comments and R code files for the compan- ion website; Scott Nestler has been a helpful friend of this book project from the beginning.

Kuber Deokar, instructional operations supervisor at Statistics.com, has been unstinting in his assistance, support, and detailed attention. We also thank Shweta Jadhav and Dhanashree Vishwasrao, assistant teachers. Valerie Troiano has shepherded many instructors and students through the Statistics.com courses that have helped nurture the development of these books.

Colleagues and family members have been providing ongoing feedback and assistance with this book project. Boaz Shmueli and Raquelle Azran gave detailed editorial comments and suggestions on the first two editions; Bruce McCullough and Adam Hughes did the same for the first edition. Noa Shmueli provided careful proofs of the third edition. Ran Shenberger offered design tips.

xxvii

xxviii ACKNOWLEDGMENTS

Ken Strasma, founder of the microtargeting firm HaystaqDNA and director of targeting for the 2004 Kerry campaign and the 2008 Obama campaign, provided the scenario and data for the section on uplift modeling. We also thank Jen Gol- beck, director of the Social Intelligence Lab at the University of Maryland and author of Analyzing the Social Web, whose book inspired our presentation in the chapter on social network analytics. Randall Pruim contributed extensively to the chapter on visualization.

Marietta Tretter at Texas A&M shared comments and thoughts on the time series chapters, and Stephen Few and Ben Shneiderman provided feedback and suggestions on the data visualization chapter and overall design tips.

Susan Palocsay and Mia Stephens have provided suggestions and feedback on numerous occasions, as has Margret Bjarnadottir. We also thank Catherine Plaisant at the University of Maryland’s Human–Computer Interaction Lab, who helped out in a major way by contributing exercises and illustrations to the data visualization chapter. Gregory Piatetsky-Shapiro, founder of KDNuggets.com, has been generous with his time and counsel over the many years of this project.

This book would not have seen the light of day without the nurturing sup- port of the faculty at the Sloan School of Management at MIT. Our special thanks to Dimitris Bertsimas, James Orlin, Robert Freund, Roy Welsch, Gor- don Kaufmann, and Gabriel Bitran. As teaching assistants for the data mining course at Sloan, Adam Mersereau gave detailed comments on the notes and cases that were the genesis of this book, Romy Shioda helped with the preparation of several cases and exercises used here, and Mahesh Kumar helped with the material on clustering.

Colleagues at the University of Maryland’s Smith School of Business: Shri- vardhan Lele, Wolfgang Jank, and Paul Zantek provided practical advice and comments. We thank Robert Windle, and University of Maryland MBA stu- dents Timothy Roach, Pablo Macouzet, and Nathan Birckhead for invaluable datasets. We also thank MBA students Rob Whitener and Daniel Curtis for the heatmap and map charts.

Anand Bodapati provided both data and advice. Jake Hofman from Micro- soft Research and Sharad Borle assisted with data access. Suresh Ankolekar and Mayank Shah helped develop several cases and provided valuable pedagogical comments. Vinni Bhandari helped write the Charles Book Club case.

We would like to thank Marvin Zelen, L. J. Wei, and Cyrus Mehta at Har- vard, as well as Anil Gore at Pune University, for thought-provoking discussions on the relationship between statistics and data mining. Our thanks to Richard Larson of the Engineering Systems Division, MIT, for sparking many stimu- lating ideas on the role of data mining in modeling complex systems. Over a decade ago, they helped us develop a balanced philosophical perspective on the emerging field of data mining.

ACKNOWLEDGMENTS xxix

Lastly, we thank the folks at Wiley for the decade-long successful journey of this book. Steve Quigley at Wiley showed confidence in this book from the beginning and helped us navigate through the publishing process with great speed. Curt Hinrichs’ vision, tips, and encouragement helped bring this book to the starting gate. Jon Gurstelle, Kathleen Pagliaro, and Katrina Maceda greatly assisted us in pushing ahead and finalizing this R edition. We are also especially grateful to Amy Hendrickson, who assisted with typesetting and making this book beautiful.

Part I

Preliminaries

CHAPTER 1

Introduction

1.1 What Is Business Analytics?

Business Analytics (BA) is the practice and art of bringing quantitative data to bear on decision-making. The term means different things to different organizations.

Consider the role of analytics in helping newspapers survive the transition to a digital world. One tabloid newspaper with a working-class readership in Britain had launched a web version of the paper, and did tests on its home page to determine which images produced more hits: cats, dogs, or monkeys. This simple application, for this company, was considered analytics. By contrast, the Washington Post has a highly influential audience that is of interest to big defense contractors: it is perhaps the only newspaper where you routinely see advertisements for aircraft carriers. In the digital environment, the Post can track readers by time of day, location, and user subscription information. In this fashion, the display of the aircraft carrier advertisement in the online paper may be focused on a very small group of individuals—say, the members of the House and Senate Armed Services Committees who will be voting on the Pentagon’s budget.

Business Analytics, or more generically, analytics, include a range of data analysis methods. Many powerful applications involve little more than count- ing, rule-checking, and basic arithmetic. For some organizations, this is what is meant by analytics.

The next level of business analytics, now termed Business Intelligence (BI), refers to data visualization and reporting for understanding “what happened and what is happening.” This is done by use of charts, tables, and dashboards to display, examine, and explore data. BI, which earlier consisted mainly of gener- ating static reports, has evolved into more user-friendly and effective tools and practices, such as creating interactive dashboards that allow the user not only to

Data Mining for Business Analytics: Concepts, Techniques, and Applications in R, First Edition. Galit Shmueli, Peter C. Bruce, Inbal Yahav, Nitin R. Patel, and Kenneth C. Lichtendahl, Jr. © 2018 John Wiley & Sons, Inc. Published 2018 by John Wiley & Sons, Inc.

3

4 INTRODUCTION

access real-time data, but also to directly interact with it. Effective dashboards are those that tie directly into company data, and give managers a tool to quickly see what might not readily be apparent in a large complex database. One such tool for industrial operations managers displays customer orders in a single two- dimensional display, using color and bubble size as added variables, showing customer name, type of product, size of order, and length of time to produce.

Business Analytics now typically includes BI as well as sophisticated data analysis methods, such as statistical models and data mining algorithms used for exploring data, quantifying and explaining relationships between measurements, and predicting new records. Methods like regression models are used to describe and quantify “on average” relationships (e.g., between advertising and sales), to predict new records (e.g., whether a new patient will react positively to a medication), and to forecast future values (e.g., next week’s web traffic).

Readers familiar with earlier editions of this book may have noticed that the book title has changed from Data Mining for Business Intelligence to Data Mining for Business Analytics in this edition. The change reflects the more recent term BA, which overtook the earlier term BI to denote advanced analytics. Today, BI is used to refer to data visualization and reporting.

W H O U S E S P R E D I C T I V E A N A L Y T I C S ?

The widespread adoption of predictive analytics, coupled with the accelerating avail- ability of data, has increased organizations’ capabilities throughout the economy. A few examples: Credit scoring: One long-established use of predictive modeling techniques for business prediction is credit scoring. A credit score is not some arbitrary judgment of credit-worthiness; it is based mainly on a predictive model that uses prior data to predict repayment behavior. Future purchases: A more recent (and controversial) example is Target’s use of predictive modeling to classify sales prospects as “pregnant” or “not-pregnant.” Those classified as pregnant could then be sent sales promotions at an early stage of pregnancy, giving Target a head start on a significant purchase stream. Tax evasion: The US Internal Revenue Service found it was 25 times more likely to find tax evasion when enforcement activity was based on predictive models, allowing agents to focus on the most-likely tax cheats (Siegel, 2013).

The Business Analytics toolkit also includes statistical experiments, the most common of which is known to marketers as A-B testing. These are often used for pricing decisions:

• Orbitz, the travel site, found that it could price hotel options higher for Mac users than Windows users.

• Staples online store found it could charge more for staplers if a customer lived far from a Staples store.

WHAT IS DATA MINING? 5

Beware the organizational setting where analytics is a solution in search of a problem: A manager, knowing that business analytics and data mining are hot areas, decides that her organization must deploy them too, to capture that hidden value that must be lurking somewhere. Successful use of analytics and data mining requires both an understanding of the business context where value is to be captured, and an understanding of exactly what the data mining methods do.

1.2 What Is Data Mining?

In this book, data mining refers to business analytics methods that go beyond counts, descriptive techniques, reporting, and methods based on business rules. While we do introduce data visualization, which is commonly the first step into more advanced analytics, the book focuses mostly on the more advanced data analytics tools. Specifically, it includes statistical and machine-learning meth- ods that inform decision-making, often in an automated fashion. Prediction is typically an important component, often at the individual level. Rather than “what is the relationship between advertising and sales,” we might be interested in “what specific advertisement, or recommended product, should be shown to a given online shopper at this moment?” Or we might be interested in clustering customers into different “personas” that receive different marketing treatment, then assigning each new prospect to one of these personas.

The era of Big Data has accelerated the use of data mining. Data mining methods, with their power and automaticity, have the ability to cope with huge amounts of data and extract value.

1.3 Data Mining and Related Terms

The field of analytics is growing rapidly, both in terms of the breadth of appli- cations, and in terms of the number of organizations using advanced analytics. As a result, there is considerable overlap and inconsistency of definitions.

The term data mining itself means different things to different people. To the general public, it may have a general, somewhat hazy and pejorative meaning of digging through vast stores of (often personal) data in search of something interesting. One major consulting firm has a “data mining department,” but its responsibilities are in the area of studying and graphing past data in search of general trends. And, to confuse matters, their more advanced predictive models are the responsibility of an “advanced analytics department.” Other terms that organizations use are predictive analytics, predictive modeling, and machine learning.

Data mining stands at the confluence of the fields of statistics and machine learning (also known as artificial intelligence). A variety of techniques for explor- ing data and building models have been around for a long time in the world of

6 INTRODUCTION

statistics: linear regression, logistic regression, discriminant analysis, and princi- pal components analysis, for example. But the core tenets of classical statistics— computing is difficult and data are scarce—do not apply in data mining applica- tions where both data and computing power are plentiful.

This gives rise to Daryl Pregibon’s description of data mining as “statistics at scale and speed” (Pregibon, 1999). Another major difference between the fields of statistics and machine learning is the focus in statistics on inference from a sample to the population regarding an “average effect”—for example, “a $1 price increase will reduce average demand by 2 boxes.” In contrast, the focus in machine learning is on predicting individual records—“the predicted demand for person i given a $1 price increase is 1 box, while for person j it is 3 boxes.” The emphasis that classical statistics places on inference (determining whether a pattern or interesting result might have happened by chance in our sample) is absent from data mining.

In comparison to statistics, data mining deals with large datasets in an open- ended fashion, making it impossible to put the strict limits around the question being addressed that inference would require. As a result, the general approach to data mining is vulnerable to the danger of overfitting, where a model is fit so closely to the available sample of data that it describes not merely structural characteristics of the data, but random peculiarities as well. In engineering terms, the model is fitting the noise, not just the signal.

In this book, we use the term machine learning to refer to algorithms that learn directly from data, especially local patterns, often in layered or iterative fashion. In contrast, we use statistical models to refer to methods that apply global structure to the data. A simple example is a linear regression model (statistical) vs. a k-nearest-neighbors algorithm (machine learning). A given record would be treated by linear regression in accord with an overall linear equation that applies to all the records. In k-nearest neighbors, that record would be classified in accord with the values of a small number of nearby records.

Lastly, many practitioners, particularly those from the IT and computer sci- ence communities, use the term machine learning to refer to all the methods dis- cussed in this book.

1.4 Big Data

Data mining and Big Data go hand in hand. Big Data is a relative term—data today are big by reference to the past, and to the methods and devices available to deal with them. The challenge Big Data presents is often characterized by the four V’s—volume, velocity, variety, and veracity. Volume refers to the amount of data. Velocity refers to the flow rate—the speed at which it is being generated and changed. Variety refers to the different types of data being generated (currency,

DATA SCIENCE 7

dates, numbers, text, etc.). Veracity refers to the fact that data is being generated by organic distributed processes (e.g., millions of people signing up for services or free downloads) and not subject to the controls or quality checks that apply to data collected for a study.

Most large organizations face both the challenge and the opportunity of Big Data because most routine data processes now generate data that can be stored and, possibly, analyzed. The scale can be visualized by comparing the data in a traditional statistical analysis (say, 15 variables and 5000 records) to the Walmart database. If you consider the traditional statistical study to be the size of a period at the end of a sentence, then the Walmart database is the size of a football field. And that probably does not include other data associated with Walmart—social media data, for example, which comes in the form of unstructured text.

If the analytical challenge is substantial, so can be the reward:

• OKCupid, the online dating site, uses statistical models with their data to predict what forms of message content are most likely to produce a response.

• Telenor, a Norwegian mobile phone service company, was able to reduce subscriber turnover 37% by using models to predict which customers were most likely to leave, and then lavishing attention on them.

• Allstate, the insurance company, tripled the accuracy of predicting injury liability in auto claims by incorporating more information about vehicle type.

The above examples are from Eric Siegel’s book Predictive Analytics (2013, Wiley). Some extremely valuable tasks were not even feasible before the era of Big

Data. Consider web searches, the technology on which Google was built. In early days, a search for “Ricky Ricardo Little Red Riding Hood” would have yielded various links to the I Love LucyTV show, other links to Ricardo’s career as a band leader, and links to the children’s story of Little Red Riding Hood. Only once the Google database had accumulated sufficient data (including records of what users clicked on) would the search yield, in the top position, links to the specific I Love Lucy episode in which Ricky enacts, in a comic mixture of Spanish and English, Little Red Riding Hood for his infant son.

1.5 Data Science

The ubiquity, size, value, and importance of Big Data has given rise to a new profession: the data scientist. Data science is a mix of skills in the areas of statistics, machine learning, math, programming, business, and IT. The term itself is thus broader than the other concepts we discussed above, and it is a rare individual who combines deep skills in all the constituent areas. In their book Analyzing

8 INTRODUCTION

the Analyzers (Harris et al., 2013), the authors describe the skill sets of most data scientists as resembling a ‘T’—deep in one area (the vertical bar of the T), and shallower in other areas (the top of the T).

At a large data science conference session (Strata+Hadoop World, Octo- ber 2014), most attendees felt that programming was an essential skill, though there was a sizable minority who felt otherwise. And, although Big Data is the motivating power behind the growth of data science, most data scientists do not actually spend most of their time working with terabyte-size or larger data.

Data of the terabyte or larger size would be involved at the deployment stage of a model. There are manifold challenges at that stage, most of them IT and programming issues related to data-handling and tying together different compo- nents of a system. Much work must precede that phase. It is that earlier piloting and prototyping phase on which this book focuses—developing the statistical and machine learning models that will eventually be plugged into a deployed system. What methods do you use with what sorts of data and problems? How do the methods work? What are their requirements, their strengths, their weaknesses? How do you assess their performance?

1.6 Why Are There So Many Different Methods?

As can be seen in this book or any other resource on data mining, there are many different methods for prediction and classification. You might ask yourself why they coexist, and whether some are better than others. The answer is that each method has advantages and disadvantages. The usefulness of a method can depend on factors such as the size of the dataset, the types of patterns that exist in the data, whether the data meet some underlying assumptions of the method, how noisy the data are, and the particular goal of the analysis. A small illustration is shown in Figure 1.1, where the goal is to find a combination of household income level and household lot size that separates buyers (solid circles) from

FIGURE 1.1 TWO METHODS FOR SEPARATING OWNERS FROM NONOWNERS

TERMINOLOGY AND NOTATION 9

nonbuyers (hollow circles) of riding mowers. The first method (left panel) looks only for horizontal and vertical lines to separate buyers from nonbuyers, whereas the second method (right panel) looks for a single diagonal line.

Different methods can lead to different results, and their performance can vary. It is therefore customary in data mining to apply several different methods and select the one that appears most useful for the goal at hand.

1.7 Terminology and Notation

Because of the hybrid parentry of data mining, its practitioners often use multiple terms to refer to the same thing. For example, in the machine learning (artificial intelligence) field, the variable being predicted is the output variable or target variable. To a statistician, it is the dependent variable or the response. Here is a summary of terms used:

Algorithm A specific procedure used to implement a particular data mining technique: classification tree, discriminant analysis, and the like.

Attribute see Predictor.

Case see Observation.

Confidence A performance measure in association rules of the type “IF A and B are purchased, THEN C is also purchased.” Confidence is the conditional probability that C will be purchased IF A and B are purchased.

Confidence also has a broader meaning in statistics (confidence interval), concern- ing the degree of error in an estimate that results from selecting one sample as opposed to another.

Dependent Variable see Response.

Estimation see Prediction.

Feature see Predictor.

Holdout Data (or holdout set) A sample of data not used in fitting a model, but instead used to assess the performance of that model. This book uses the terms validation set and test set instead of holdout set.

Input Variable see Predictor.

Model An algorithm as applied to a dataset, complete with its settings (many of the algorithms have parameters that the user can adjust).

Observation The unit of analysis on which the measurements are taken (a cus- tomer, a transaction, etc.), also called instance, sample, example, case, record, pattern, or row. In spreadsheets, each row typically represents a record; each column, a variable. Note that the use of the term “sample” here is dif- ferent from its usual meaning in statistics, where it refers to a collection of observations.

10 INTRODUCTION

Outcome Variable see Response.

Output Variable see Response.

P (A | B) The conditional probability of event A occurring given that event B has occurred, read as “the probability that A will occur given that B has occurred.”

Prediction The prediction of the numerical value of a continuous output vari- able; also called estimation.

Predictor A variable, usually denoted by X , used as an input into a predic- tive model, also called a feature, input variable, independent variable, or from a database perspective, a field.

Profile A set of measurements on an observation (e.g., the height, weight, and age of a person).

Record see Observation.

Response A variable, usually denoted by Y , which is the variable being pre- dicted in supervised learning, also called dependent variable, output variable, target variable, or outcome variable.

Sample In the statistical community, “sample” means a collection of observa- tions. In the machine learning community, “sample” means a single obser- vation.

Score A predicted value or class. Scoring new data means using a model devel- oped with training data to predict output values in new data.

Success Class The class of interest in a binary outcome (e.g., purchasers in the outcome purchase/no purchase).

Supervised Learning The process of providing an algorithm (logistic regres- sion, regression tree, etc.) with records in which an output variable of inter- est is known and the algorithm “learns” how to predict this value with new records where the output is unknown.

Target see Response.

Test Data (or test set) The portion of the data used only at the end of the model building and selection process to assess how well the final model might perform on new data.

Training Data (or training set) The portion of the data used to fit a model.

Unsupervised Learning An analysis in which one attempts to learn patterns in the data other than predicting an output value of interest.

Validation Data (or validation set) The portion of the data used to assess how well the model fits, to adjust models, and to select the best model from among those that have been tried.

Variable Any measurement on the records, including both the input (X) vari- ables and the output (Y ) variable.

ROAD MAPS TO THIS BOOK 11

1.8 Road Maps to This Book

The book covers many of the widely used predictive and classification methods as well as other data mining tools. Figure 1.2 outlines data mining from a process perspective and where the topics in this book fit in. Chapter numbers are indi- cated beside the topic. Table 1.1 provides a different perspective: it organizes data mining procedures according to the type and structure of the data.

Order of Topics

The book is divided into five parts: Part I (Chapters 1–2) gives a general overview of data mining and its components. Part II (Chapters 3–4) focuses on the early stages of data exploration and dimension reduction.

Part III (Chapter 5) discusses performance evaluation. Although it contains only one chapter, we discuss a variety of topics, from predictive performance metrics to misclassification costs. The principles covered in this part are crucial for the proper evaluation and comparison of supervised learning methods.

Part IV includes eight chapters (Chapters 6–13), covering a variety of popular supervised learning methods (for classification and/or prediction). Within this part, the topics are generally organized according to the level of sophistication of the algorithms, their popularity, and ease of understanding. The final chapter introduces ensembles and combinations of methods.

FIGURE 1.2 DATA MINING FROM A PROCESS PERSPECTIVE. NUMBERS IN PARENTHESES INDICATE CHAPTER NUMBERS

12 INTRODUCTION

TABLE 1.1 ORGANIZATION OF DATA MINING METHODS IN THIS BOOK, ACCORDING TO THE NATURE OF THE DATA∗

Supervised Unsupervised Continuous Categorical Response Response No Response

Continuous Linear regression (6) Logistic regression (10) Principal components (4) predictors Neural nets (11) Neural nets (11) Cluster analysis (15)

k-Nearest neighbors (7) Discriminant analysis (12) Collaborative filtering (14)

Ensembles (13) k-Nearest neighbors (7) Ensembles (13)

Categorical Linear regression (6) Neural nets (11) Association rules (14) predictors Neural nets (11) Classification trees (9) Collaborative filtering (14)

Regression trees (9) Logistic regression (10) Ensembles (13) Naive Bayes (8)

Ensembles (13)

∗Numbers in parentheses indicate chapter number.

Part V focuses on unsupervised mining of relationships. It presents associa- tion rules and collaborative filtering (Chapter 14) and cluster analysis (Chapter 15).

Part VI includes three chapters (Chapters 16–18), with the focus on fore- casting time series. The first chapter covers general issues related to handling and understanding time series. The next two chapters present two popular fore- casting approaches: regression-based forecasting and smoothing methods.

Part VII (Chapters 19–20) presents two broad data analytics topics: social network analysis and text mining. These methods apply data mining to special- ized data structures: social networks and text.

Finally, part VIII includes a set of cases. Although the topics in the book can be covered in the order of the chapters,

each chapter stands alone. We advise, however, to read parts I–III before pro- ceeding to chapters in parts IV–V. Similarly, Chapter 16 should precede other chapters in part VI.

U S I N G R A N D R S T U D I O

To facilitate a hands-on data mining experience, this book uses R, a free software environment for statistical computing and graphics, and RStudio, an integrated development environment (IDE) for R. The R programming language is widely used in academia and industry for data mining and data analysis. R offers a variety of methods for analyzing data, provided by a variety of separate packages. Among the numerous packages, R has extensive coverage of statistical and data mining tech- niques for classification, prediction, mining associations and text, forecasting, and

ROAD MAPS TO THIS BOOK 13

data exploration and reduction. It offers a variety of supervised data mining tools: neural nets, classification and regression trees, k-nearest-neighbor classification, naive Bayes, logistic regression, linear regression, and discriminant analysis, all for predictive modeling. R’s packages also cover unsupervised algorithms: association rules, collaborative filtering, principal components analysis, k-means clustering, and hierarchical clustering, as well as visualization tools and data-handling util- ities. Often, the same method is implemented in multiple packages, as we will discuss throughout the book. The illustrations, exercises, and cases in this book are written in relation to R.

Download: To download R and RStudio, visit www.r-project.org and www.rstudio. com/products/RStudio and follow the instructions there.

Installation: Install both R and RStudio. Note that R releases new versions fairly often. When a new version is released, some packages might require a new instal- lation of R (this is rare).

Use: To start using R, open RStudio, then open a new script under File > New File > R Script. RStudio contains four panels as shown in Figure 1.3: Script (top left), Console (bottom left), Environment (top right), and additional information, such as plot and help (bottom right). To run a selected code line from the Script panel, press ctrl+r. Code lines starting with # are comments.

Package Installation: To start using an R package, you will first need to install it. Installation is done via the information panel (tab ”packages”) or using command install.packages(). New packages might not support old R versions and require a new R installation.

FIGURE 1.3 RSTUDIO SCREEN

CHAPTER 2

Overview of the Data Mining Process

In this chapter, we give an overview of the steps involved in data mining, starting from a clear goal definition and ending with model deployment. The general steps are shown schematically in Figure 2.1. We also discuss issues related to data collection, cleaning, and preprocessing. We introduce the notion of data partitioning, where methods are trained on a set of training data and then their performance is evaluated on a separate set of validation data, as well as explain how this practice helps avoid overfitting. Finally, we illustrate the steps of model building by applying them to data.

Define Purpose

Obtain Data

Explore & Clean

Data

Determine DM Task

Choose DM

Methods

Apply Methods & Select Final

Model

Evaluate Performance

Deploy

FIGURE 2.1 SCHEMATIC OF THE DATA MODELING PROCESS

2.1 Introduction

In Chapter 1, we saw some very general definitions of data mining. In this chap- ter, we introduce the variety of methods sometimes referred to as data mining. The core of this book focuses on what has come to be called predictive analyt- ics, the tasks of classification and prediction as well as pattern discovery, which have become key elements of a “business analytics” function in most large firms. These terms are described and illustrated below.

Data Mining for Business Analytics: Concepts, Techniques, and Applications in R, First Edition. Galit Shmueli, Peter C. Bruce, Inbal Yahav, Nitin R. Patel, and Kenneth C. Lichtendahl, Jr. © 2018 John Wiley & Sons, Inc. Published 2018 by John Wiley & Sons, Inc.

15

16 OVERVIEW OF THE DATA MINING PROCESS

Not covered in this book to any great extent are two simpler database meth- ods that are sometimes considered to be data mining techniques: (1) OLAP (online analytical processing) and (2) SQL (structured query language). OLAP and SQL searches on databases are descriptive in nature and are based on business rules set by the user (e.g., “find all credit card customers in a certain zip code with annual charges > $20,000, who own their home and who pay the entire amount of their monthly bill at least 95% of the time.”) Although SQL queries are often used to obtain the data in data mining, they do not involve statistical modeling or automated algorithmic methods.

2.2 Core Ideas in Data Mining

Classification

Classification is perhaps the most basic form of data analysis. The recipient of an offer can respond or not respond. An applicant for a loan can repay on time, repay late, or declare bankruptcy. A credit card transaction can be normal or fraudulent. A packet of data traveling on a network can be benign or threatening. A bus in a fleet can be available for service or unavailable. The victim of an illness can be recovered, still be ill, or be deceased.

A common task in data mining is to examine data where the classification is unknown or will occur in the future, with the goal of predicting what that classification is or will be. Similar data where the classification is known are used to develop rules, which are then applied to the data with the unknown classification.

Prediction

Prediction is similar to classification, except that we are trying to predict the value of a numerical variable (e.g., amount of purchase) rather than a class (e.g., purchaser or nonpurchaser). Of course, in classification we are trying to predict a class, but the term prediction in this book refers to the prediction of the value of a continuous variable. (Sometimes in the data mining literature, the terms estima- tion and regression are used to refer to the prediction of the value of a continuous variable, and prediction may be used for both continuous and categorical data.)

Association Rules and Recommendation Systems

Large databases of customer transactions lend themselves naturally to the analysis of associations among items purchased, or “what goes with what.” Association rules, or affinity analysis, is designed to find such general associations patterns between items in large databases. The rules can then be used in a variety of ways. For example, grocery stores can use such information for product placement.

CORE IDEAS IN DATA MINING 17

They can use the rules for weekly promotional offers or for bundling products. Association rules derived from a hospital database on patients’ symptoms during consecutive hospitalizations can help find “which symptom is followed by what other symptom” to help predict future symptoms for returning patients.

Online recommendation systems, such as those used on Amazon.com and Netflix.com, use collaborative filtering, a method that uses individual users’ pref- erences and tastes given their historic purchase, rating, browsing, or any other measurable behavior indicative of preference, as well as other users’ history. In contrast to association rules that generate rules general to an entire population, collaborative filtering generates “what goes with what” at the individual user level. Hence, collaborative filtering is used in many recommendation systems that aim to deliver personalized recommendations to users with a wide range of preferences.

Predictive Analytics

Classification, prediction, and to some extent, association rules and collaborative filtering constitute the analytical methods employed in predictive analytics. The term predictive analytics is sometimes used to also include data pattern identifi- cation methods such as clustering.

Data Reduction and Dimension Reduction

The performance of data mining algorithms is often improved when the num- ber of variables is limited, and when large numbers of records can be grouped into homogeneous groups. For example, rather than dealing with thousands of product types, an analyst might wish to group them into a smaller number of groups and build separate models for each group. Or a marketer might want to classify customers into different “personas,” and must therefore group customers into homogeneous groups to define the personas. This process of consolidating a large number of records (or cases) into a smaller set is termed data reduction. Methods for reducing the number of cases are often called clustering.

Reducing the number of variables is typically called dimension reduction. Dimension reduction is a common initial step before deploying data min- ing methods, intended to improve predictive power, manageability, and inter- pretability.

Data Exploration and Visualization

One of the earliest stages of engaging with a dataset is exploring it. Exploration is aimed at understanding the global landscape of the data, and detecting unusual values. Exploration is used for data cleaning and manipulation as well as for visual discovery and “hypothesis generation.”

18 OVERVIEW OF THE DATA MINING PROCESS

Methods for exploring data include looking at various data aggregations and summaries, both numerically and graphically. This includes looking at each variable separately as well as looking at relationships among variables. The pur- pose is to discover patterns and exceptions. Exploration by creating charts and dashboards is called Data Visualization or Visual Analytics. For numerical vari- ables, we use histograms and boxplots to learn about the distribution of their values, to detect outliers (extreme observations), and to find other information that is relevant to the analysis task. Similarly, for categorical variables, we use bar charts. We can also look at scatter plots of pairs of numerical variables to learn about possible relationships, the type of relationship, and again, to detect outliers. Visualization can be greatly enhanced by adding features such as color and interactive navigation.

Supervised and Unsupervised Learning

A fundamental distinction among data mining techniques is between supervised and unsupervised methods. Supervised learning algorithms are those used in classi- fication and prediction. We must have data available in which the value of the outcome of interest (e.g., purchase or no purchase) is known. Such data are also called “labeled data,” since they contain the label (outcome value) for each record. These training data are the data from which the classification or predic- tion algorithm “learns,” or is “trained,” about the relationship between predictor variables and the outcome variable. Once the algorithm has learned from the training data, it is then applied to another sample of labeled data (the validation data) where the outcome is known but initially hidden, to see how well it does in comparison to other models. If many different models are being tried out, it is prudent to save a third sample, which also includes known outcomes (the test data) to use with the model finally selected to predict how well it will do. The model can then be used to classify or predict the outcome of interest in new cases where the outcome is unknown.

Simple linear regression is an example of a supervised learning algorithm (although rarely called that in the introductory statistics course where you prob- ably first encountered it). The Y variable is the (known) outcome variable and the X variable is a predictor variable. A regression line is drawn to minimize the sum of squared deviations between the actual Y values and the values predicted by this line. The regression line can now be used to predict Y values for new values of X for which we do not know the Y value.

Unsupervised learning algorithms are those used where there is no outcome variable to predict or classify. Hence, there is no “learning” from cases where such an outcome variable is known. Association rules, dimension reduction methods, and clustering techniques are all unsupervised learning methods.

THE STEPS IN DATA MINING 19

Supervised and unsupervised methods are sometimes used in conjunction. For example, unsupervised clustering methods are used to separate loan appli- cants into several risk-level groups. Then, supervised algorithms are applied separately to each risk-level group for predicting propensity of loan default.

S U P E R V I S E D L E A R N I N G R E Q U I R E S G O O D S U P E R V I S I O N

In some cases, the value of the outcome variable (the ‘label’) is known because it is an inherent component of the data. Web logs will show whether a person clicked on a link or not. Bank records will show whether a loan was paid on time or not. In other cases, the value of the known outcome must be supplied by a human labeling process to accumulate enough data to train a model. E-mail must be labeled as spam or legitimate, documents in legal discovery must be labeled as relevant or irrelevant. In either case, the data mining algorithm can be led astray if the quality of the supervision is poor.

Gene Weingarten reported in the January 5, 2014 Washington Post magazine how the strange phrase “defiantly recommend” is making its way into English via auto-correction. “Defiantly” is closer to the common misspelling definatly than is definitely, so Google.com, in the early days, offered it as a correction when users typed the misspelled word “definatly.” In the ideal supervised learning model, humans guide the auto-correction process by rejecting defiantly and substituting definitely. Google’s algorithm would then learn that this is the best first-choice correction of “definatly.” The problem was that too many people were lazy, just accepting the first correction that Google presented. All these acceptances then cemented “defiantly” as the proper correction.

2.3 The Steps in Data Mining

This book focuses on understanding and using data mining algorithms (Steps 4 to 7 below). However, some of the most serious errors in analytics projects result from a poor understanding of the problem—an understanding that must be developed before we get into the details of algorithms to be used. Here is a list of steps to be taken in a typical data mining effort:

1. Develop an understanding of the purpose of the data mining project. How will the stakeholder use the results? Who will be affected by the results? Will the analysis be a one-shot effort or an ongoing procedure?

2. Obtain the dataset to be used in the analysis. This often involves sampling from a large database to capture records to be used in an analysis. How well this sample reflects the records of interest affects the ability of the data mining results to generalize to records outside of this sample. It may also involve pulling together data from different databases or sources.

20 OVERVIEW OF THE DATA MINING PROCESS

The databases could be internal (e.g., past purchases made by customers) or external (credit ratings). While data mining deals with very large databases, usually the analysis to be done requires only thousands or tens of thousands of records.

3. Explore, clean, and preprocess the data. This step involves verifying that the data are in reasonable condition. How should missing data be handled? Are the values in a reasonable range, given what you would expect for each variable? Are there obvious outliers? The data are reviewed graph- ically: for example, a matrix of scatterplots showing the relationship of each variable with every other variable. We also need to ensure con- sistency in the definitions of fields, units of measurement, time periods, and so on. In this step, new variables are also typically created from exist- ing ones. For example, “duration” can be computed from start and end dates.

4. Reduce the data dimension, if necessary. Dimension reduction can involve operations such as eliminating unneeded variables, transforming variables (e.g., turning “money spent” into “spent > $100” vs. “spent ≤ $100”), and creating new variables (e.g., a variable that records whether at least one of several products was purchased). Make sure that you know what each variable means and whether it is sensible to include it in the model.

5. Determine the data mining task. (classification, prediction, clustering, etc.). This involves translating the general question or problem of Step 1 into a more specific data mining question.

6. Partition the data (for supervised tasks). If the task is supervised (classification or prediction), randomly partition the dataset into three parts: training, validation, and test datasets.

7. Choose the data mining techniques to be used. (regression, neural nets, hier- archical clustering, etc.).

8. Use algorithms to perform the task. This is typically an iterative process— trying multiple variants, and often using multiple variants of the same algorithm (choosing different variables or settings within the algorithm). Where appropriate, feedback from the algorithm’s performance on vali- dation data is used to refine the settings.

9. Interpret the results of the algorithms. This involves making a choice as to the best algorithm to deploy, and where possible, testing the final choice on the test data to get an idea as to how well it will perform. (Recall that each algorithm may also be tested on the validation data for tuning purposes; in this way, the validation data become a part of the fitting process and are likely to underestimate the error in the deployment of the model that is finally chosen.)

PRELIMINARY STEPS 21

10. Deploy the model. This step involves integrating the model into oper- ational systems and running it on real records to produce decisions or actions. For example, the model might be applied to a purchased list of possible customers, and the action might be “include in the mailing if the predicted amount of purchase is > $10.” A key step here is “scoring” the new records, or using the chosen model to predict the outcome value (“score”) for each new record.

The foregoing steps encompass the steps in SEMMA, a methodology developed by the software company SAS:

Sample Take a sample from the dataset; partition into training, validation, and test datasets.

Explore Examine the dataset statistically and graphically.

Modify Transform the variables and impute missing values.

Model Fit predictive models (e.g., regression tree, neural network).

Assess Compare models using a validation dataset.

IBM SPSS Modeler (previously SPSS-Clementine) has a similar method- ology, termed CRISP-DM (CRoss-Industry Standard Process for Data Min- ing). All these frameworks include the same main steps involved in predictive modeling.

2.4 Preliminary Steps

Organization of Datasets

Datasets are nearly always constructed and displayed so that variables are in columns and records are in rows. We will illustrate this with home values in West Roxbury, Boston, in 2014. 14 variables are recorded for over 5000 homes. The spreadsheet is organized so that each row represents a home—the first home’s assessed value was $344,200, its tax was $4430, its size was 9965 ft2, it was built in 1880, and so on. In supervised learning situations, one of these variables will be the outcome variable, typically listed in the first or last column (in this case it is TOTAL VALUE, in the first column).

Predicting Home Values in the West Roxbury Neighborhood

The Internet has revolutionized the real estate industry. Realtors now list houses and their prices on the web, and estimates of house and condominium prices have become widely available, even for units not on the market. At this time of

22 OVERVIEW OF THE DATA MINING PROCESS

writing, Zillow (www.zillow.com) is the most popular online real estate infor- mation site in the United States1, and in 2014 they purchased their major rival, Trulia. By 2015, Zillow had become the dominant platform for checking house prices and, as such, the dominant online advertising venue for realtors. What used to be a comfortable 6% commission structure for realtors, affording them a handsome surplus (and an oversupply of realtors), was being rapidly eroded by an increasing need to pay for advertising on Zillow. (This, in fact, is the key to Zillow’s business model—redirecting the 6% commission away from realtors and to itself.)

Zillow gets much of the data for its “Zestimates” of home values directly from publicly available city housing data, used to estimate property values for tax assessment. A competitor seeking to get into the market would likely take the same approach. So might realtors seeking to develop an alternative to Zillow.

A simple approach would be a naive, model-less method—just use the assessed values as determined by the city. Those values, however, do not nec- essarily include all properties, and they might not include changes warranted by remodeling, additions, etc. Moreover, the assessment methods used by cities may not be transparent or always reflect true market values. However, the city property data can be used as a starting point to build a model, to which additional data (such as that collected by large realtors) can be added later.

Let’s look at how Boston property assessment data, available from the city of Boston, might be used to predict home values. The data in WestRoxbury.csv includes information on single family owner-occupied homes in West Roxbury, a neighborhood in southwest Boston, MA, in 2014. The data include values for various predictor variables, and for an outcome—assessed home value (“total value”). This dataset has 14 variables, and a description of each variable is given in Table 2.1 (the full data dictionary provided by the City of Boston is available at http://goo.gl/QBRlYF; we have modified a few variable names). The dataset includes 5802 homes. A sample of the data is shown in Table 2.2, and the “data dictionary” describing each variable is in Table 2.1.

As we saw earlier, below the header row, each row in the data represents a home. For example, the first home was assessed at a total value of $344.2 thousand (TOTAL VALUE). Its tax bill was $4330. It has a lot size of 9965 square feet (ft2), was built in the year 1880, has two floors, six rooms, and so on.

Loading and Looking at the Data in R

To load data into R, we will typically want to have the data available as a csv (comma separated values) file. If the data are in an xls (or xlsx) file, we can save

1“Harney, K., Zestimates may not be as right as you’d like”, Washington Post, Feb. 7, 2015, p. T10.

PRELIMINARY STEPS 23

TABLE 2.1 DESCRIPTION OF VARIABLES IN WEST ROXBURY (BOSTON) HOME VALUE DATASET

TOTAL VALUE Total assessed value for property, in thousands of USD

TAX Tax bill amount based on total assessed value multiplied by the tax rate, in USD

LOT SQ FT Total lot size of parcel in square feet

YR BUILT Year the property was built

GROSS AREA Gross floor area

LIVING AREA Total living area for residential properties (ft2)

FLOORS Number of floors

ROOMS Total number of rooms

BEDROOMS Total number of bedrooms

FULL BATH Total number of full baths

HALF BATH Total number of half baths

KITCHEN Total number of kitchens

FIREPLACE Total number of fireplaces

REMODEL When the house was remodeled (Recent/Old/None)

that same file in Excel as a csv file: go to File > Save as > Save as type: CSV (Comma delimited) (*.csv) > Save. Note: When dealing with .csv files in Excel, beware of two things:

• Opening a .csv file in Excel strips off leading 0’s, which corrupts zipcode data.

• Saving a .csv file in Excel saves only the digits that are displayed; if you need precision to a certain number of decimals, you need to ensure they are displayed before saving.

Once we have R and RStudio installed on our machine and the West Rox- bury.csv file saved as a csv file, we can run the code in Table 2.3 to load the data into R.

TABLE 2.2 FIRST 10 RECORDS IN THE WEST ROXBURY HOME VALUES DATASET

TOTAL TAX LOT YR GROSS LIVING FLOORS ROOMS BED FULL HALF KIT FIRE REMODEL

VALUE SQ FT BUILT AREA AREA ROOMS BATH BATH CHEN PLACE

344.2 4330 9965 1880 2436 1352 2 6 3 1 1 1 0 None

412.6 5190 6590 1945 3108 1976 2 10 4 2 1 1 0 Recent

330.1 4152 7500 1890 2294 1371 2 8 4 1 1 1 0 None

498.6 6272 13,773 1957 5032 2608 1 9 5 1 1 1 1 None

331.5 4170 5000 1910 2370 1438 2 7 3 2 0 1 0 None

337.4 4244 5142 1950 2124 1060 1 6 3 1 0 1 1 Old

359.4 4521 5000 1954 3220 1916 2 7 3 1 1 1 0 None

320.4 4030 10,000 1950 2208 1200 1 6 3 1 0 1 0 None

333.5 4195 6835 1958 2582 1092 1 5 3 1 0 1 1 Recent

409.4 5150 5093 1900 4818 2992 2 8 4 2 0 1 0 None

24 OVERVIEW OF THE DATA MINING PROCESS

TABLE 2.3 WORKING WITH FILES IN R

To start, open RStudio, go to File > New File > R Script. It opens a new tab. Then save your Untitled1.R file into the directory where your csv is saved. Give it the name WestRoxbury.R. From the Menu Bar, go to Session > Set Working Directory > To Source File Location; This sets the working directory as the place where both the R file and csv file are saved.

code for loading and creating subsets from the data

housing.df <- read.csv("WestRoxbury.csv", header = TRUE) # load data dim(housing.df) # find the dimension of data frame head(housing.df) # show the first six rows View(housing.df) # show all the data in a new tab

# Practice showing different subsets of the data housing.df[1:10, 1] # show the first 10 rows of the first column only housing.df[1:10, ] # show the first 10 rows of each of the columns housing.df[5, 1:10] # show the fifth row of the first 10 columns housing.df[5, c(1:2, 4, 8:10)] # show the fifth row of some columns housing.df[, 1] # show the whole first column housing.df$TOTAL_VALUE # a different way to show the whole first column housing.df$TOTAL_VALUE[1:10] # show the first 10 rows of the first column length(housing.df$TOTAL_VALUE) # find the length of the first column mean(housing.df$TOTAL_VALUE) # find the mean of the first column summary(housing.df) # find summary statistics for each column

Data from a csv file is stored in R as a data frame (e.g., housing.df ). If our csv file has column headers, these headers get automatically stored as the column names of our data. A data frame is the fundamental object almost all analyses begin with in R. A data frame has rows and columns. The rows are the obser- vations for each case (e.g., house), and the columns are the variables of interest (e.g., TOTAL VALUE, TAX). The code in Table 2.3 walks you through some basic steps you will want to perform prior to doing any analysis: finding the size and dimension of your data (number of rows and columns), viewing all the data, displaying only selected rows and columns, and computing summary statistics for variables of interest. Note that comments are preceded with the # symbol.

Sampling from a Database

Typically, we perform data mining on less than the complete database. Data mining algorithms will have varying limitations on what they can handle in terms of the numbers of records and variables, limitations that may be specific to computing power and capacity as well as software limitations. Even within those limits, many algorithms will execute faster with smaller samples.

Accurate models can often be built with as few as several thousand records. Hence, we will want to sample a subset of records for model building. Table 2.4 provides code for sampling in R.

PRELIMINARY STEPS 25

TABLE 2.4 SAMPLING IN R

code for sampling and over/under-sampling

# random sample of 5 observations s <- sample(row.names(housing.df), 5) housing.df[s,]

# oversample houses with over 10 rooms s <- sample(row.names(housing.df), 5, prob = ifelse(housing.df$ROOMS>10, 0.9, 0.01)) housing.df[s,]

Oversampling Rare Events in Classification Tasks

If the event we are interested in classifying is rare, for example, customers pur- chasing a product in response to a mailing, or fraudulent credit card transactions, sampling a random subset of records may yield so few events (e.g., purchases) that we have little information on them. We would end up with lots of data on nonpurchasers and non-fraudulent transactions but little on which to base a model that distinguishes purchasers from nonpurchasers or fraudulent from non- fraudulent. In such cases, we would want our sampling procedure to overweight the rare class (purchasers or frauds) relative to the majority class (nonpurchasers, non-frauds) so that our sample would end up with a healthy complement of purchasers or frauds.

Assuring an adequate number of responder or “success” cases to train the model is just part of the picture. A more important factor is the costs of mis- classification. Whenever the response rate is extremely low, we are likely to attach more importance to identifying a responder than to identifying a non- responder. In direct-response advertising (whether by traditional mail, e-mail, or web advertising), we may encounter only one or two responders for every hundred records—the value of finding such a customer far outweighs the costs of reaching him or her. In trying to identify fraudulent transactions, or customers unlikely to repay debt, the costs of failing to find the fraud or the nonpaying customer are likely to exceed the cost of more detailed review of a legitimate transaction or customer.

If the costs of failing to locate responders are comparable to the costs of misidentifying responders as non-responders, our models would usually achieve highest overall accuracy if they identified everyone as a non-responder (or almost everyone, if it is easy to identify a few responders without catching many non- responders). In such a case, the misclassification rate is very low—equal to the rate of responders—but the model is of no value.

26 OVERVIEW OF THE DATA MINING PROCESS

More generally, we want to train our model with the asymmetric costs in mind so that the algorithm will catch the more valuable responders, probably at the cost of “catching” and misclassifying more non-responders as responders than would be the case if we assume equal costs. This subject is discussed in detail in Chapter 5.

Preprocessing and Cleaning the Data

Types of Variables There are several ways of classifying variables. Vari- ables can be numerical or text (character/string). They can be continuous (able to assume any real numerical value, usually in a given range), integer (taking only integer values), categorical (assuming one of a limited number of values), or date. Categorical variables can be either coded as numerical (1, 2, 3) or text (payments current, payments not current, bankrupt). Categorical variables can be unordered (called nominal variables) with categories such as North America, Europe, and Asia; or they can be ordered (called ordinal variables) with categories such as high value, low value, and nil value.

Continuous variables can be handled by most data mining routines with the exception of the naive Bayes classifier, which deals exclusively with categorical predictor variables. The machine learning roots of data mining grew out of problems with categorical outcomes; the roots of statistics lie in the analysis of continuous variables. Sometimes, it is desirable to convert continuous variables to categorical variables. This is done most typically in the case of outcome variables, where the numerical variable is mapped to a decision (e.g., credit scores above a certain threshold mean “grant credit,” a medical test result above a certain threshold means “start treatment”).

For the West Roxbury data, Table 2.5 presents some R statements to review the variables and determine what type (class) R thinks they are, and to determine the number of levels in a factor variable.

Handling Categorical Variables Categorical variables can also be han- dled by most data mining routines, but often require special handling. If the categorical variable is ordered (age group, degree of creditworthiness, etc.), we can sometimes code the categories numerically (1, 2, 3, ...) and treat the vari- able as if it were a continuous variable. The smaller the number of categories, and the less they represent equal increments of value, the more problematic this approach becomes, but it often works well enough.

Nominal categorical variables, however, often cannot be used as is. In many cases, they must be decomposed into a series of binary variables, called dummy

PRELIMINARY STEPS 27

TABLE 2.5 REVIEWING VARIABLES IN R

code for reviewing variables

names(housing.df) # print a list of variables to the screen. t(t(names(housing.df))) # print the list in a useful column format colnames(housing.df)[1] <- c("TOTAL_VALUE") # change the first column's name class(housing.df$REMODEL) # REMODEL is a factor variable class(housing.df[ ,14]) # Same. levels(housing.df[, 14]) # It can take one of three levels class(housing.df$BEDROOMS) # BEDROOMS is an integer variable class(housing.df[, 1]) # Total_Value is a numeric variable

Partial Output

> t(t(names(housing.df))) [,1]

[1,] "TOTAL_VALUE" [2,] "TAX" [3,] "LOT.SQFT" [4,] "YR.BUILT" [5,] "GROSS.AREA" [6,] "LIVING.AREA" [7,] "FLOORS" [8,] "ROOMS" [9,] "BEDROOMS" [10,] "FULL.BATH" [11,] "HALF.BATH" [12,] "KITCHEN" [13,] "FIREPLACE" [14,] "REMODEL"

> class(housing.df$REMODEL) [1] "factor"

> levels(housing.df[, 14]) [1] "None" "Old" "Recent"

variables. For example, a single categorical variable that can have possible values of “student,” “unemployed,” “employed,” or “retired” would be split into four separate dummy variables:

Student—Yes/No Unemployed—Yes/No Employed—Yes/No Retired—Yes/No

In many cases, only three of the dummy variables need to be used; if the val- ues of three are known, the fourth is also known. For example, given that these four values are the only possible ones, we can know that if a person is neither

28 OVERVIEW OF THE DATA MINING PROCESS

student, unemployed, nor employed, he or she must be retired. In some rou- tines (e.g., linear regression and logistic regression), you should not use all four variables—the redundant information will cause the algorithm to fail. Note, also, that typical methods of creating dummy variables will leave the original categorical variable intact; obviously you should not use both the original vari- able and the dummies. The R code to create binary dummies from a categorical (factor) variable is given in Table 2.6.

TABLE 2.6 CREATING DUMMY VARIABLES IN R

code for creating binary dummies (indicators)

# use model.matrix() to convert all categorical variables in the data frame into # a set of dummy variables. We must then turn the resulting data matrix back into # a data frame for further work. xtotal <- model.matrix(~ 0 + BEDROOMS + REMODEL, data = housing.df) xtotal$BEDROOMS[1:5] # will not work because xtotal is a matrix xtotal <- as.data.frame(xtotal) t(t(names(xtotal))) # check the names of the dummy variables head(xtotal) xtotal <- xtotal[, -4] # drop one of the dummy variables. # In this case, drop REMODELRecent.

Partial Output

> t(t(names(xtotal))) # Check the names of the dummy variables. [,1]

[1,] "BEDROOMS" [2,] "REMODELNone" [3,] "REMODELOld" [4,] "REMODELRecent"

> head(xtotal) BEDROOMS REMODELNone REMODELOld REMODELRecent

1 3 1 0 0 2 4 0 0 1 3 4 1 0 0 4 5 1 0 0 5 3 1 0 0 6 3 0 1 0

Variable Selection More is not necessarily better when it comes to select- ing variables for a model. Other things being equal, parsimony, or compactness, is a desirable feature in a model. For one thing, the more variables we include and the more complex the model, the greater the number of records we will need to assess relationships among the variables. Fifteen records may suffice to give us a rough idea of the relationship between Y and a single predictor vari- able X . If we now want information about the relationship between Y and 15 predictor variables X1, . . . , X15, 15 records will not be enough (each estimated

PRELIMINARY STEPS 29

relationship would have an average of only one record’s worth of information, making the estimate very unreliable). In addition, models based on many vari- ables are often less robust, as they require the collection of more variables in the future, are subject to more data quality and availability issues, and require more data cleaning and preprocessing.

How Many Variables and How Much Data? Statisticians give us proce- dures to learn with some precision how many records we would need to achieve a given degree of reliability with a given dataset and a given model. These are called “power calculations” and are intended to assure that an average pop- ulation effect will be estimated with sufficient precision from a sample. Data miners’ needs are usually different, because the focus is not on identifying an average effect but rather on predicting individual records. This purpose typically requires larger samples than those used for statistical inference. A good rule of thumb is to have 10 records for every predictor variable. Another rule, used by Delmaster and Hancock (2001, p. 68) for classification procedures, is to have at least 6×m× p records, where m is the number of outcome classes and p is the number of variables.

In general, compactness or parsimony is a desirable feature in a data mining model. Even when we start with a small number of variables, we often end up with many more after creating new variables (such as converting a categor- ical variable into a set of dummy variables). Data visualization and dimension reduction methods help reduce the number of variables so that redundancies and information overlap are reduced.

Even when we have an ample supply of data, there are good reasons to pay close attention to the variables that are included in a model. Someone with domain knowledge (i.e., knowledge of the business process and the data) should be consulted, as knowledge of what the variables represent is typically critical for building a good model and avoiding errors. For example, suppose we’re trying to predict the total purchase amount spent by customers, and we have a few predictor columns that are coded X1, X2, X3, . . ., where we don’t know what those codes mean. We might find that X1 is an excellent predictor of the total amount spent. However, if we discover that X1 is the amount spent on shipping, calculated as a percentage of the purchase amount, then obviously a model that uses shipping amount cannot be used to predict purchase amount, because the shipping amount is not known until the transaction is completed. Another example is if we are trying to predict loan default at the time a customer applies for a loan. If our dataset includes only information on approved loan applications, we will not have information about what distinguishes defaulters from non-defaulters among denied applicants. A model based on approved loans alone can therefore not be used to predict defaulting behavior at the time of loan application, but rather only once a loan is approved.

30 OVERVIEW OF THE DATA MINING PROCESS

Outliers The more data we are dealing with, the greater the chance of encountering erroneous values resulting from measurement error, data-entry error, or the like. If the erroneous value is in the same range as the rest of the data, it may be harmless. If it is well outside the range of the rest of the data (a misplaced decimal, for example), it may have a substantial effect on some of the data mining procedures we plan to use.

Values that lie far away from the bulk of the data are called outliers. The term far away is deliberately left vague because what is or is not called an outlier is an arbitrary decision. Analysts use rules of thumb such as “anything over three standard deviations away from the mean is an outlier,” but no statistical rule can tell us whether such an outlier is the result of an error. In this statistical sense, an outlier is not necessarily an invalid data point, it is just a distant one.

The purpose of identifying outliers is usually to call attention to values that need further review. We might come up with an explanation looking at the data—in the case of a misplaced decimal, this is likely. We might have no expla- nation, but know that the value is wrong—a temperature of 178◦F for a sick person. Or, we might conclude that the value is within the realm of possibility and leave it alone. All these are judgments best made by someone with domain knowledge, knowledge of the particular application being considered: direct mail, mortgage finance, and so on, as opposed to technical knowledge of statistical or data mining procedures. Statistical procedures can do little beyond identifying the record as something that needs review.

If manual review is feasible, some outliers may be identified and corrected. In any case, if the number of records with outliers is very small, they might be treated as missing data. How do we inspect for outliers? One technique is to sort the records by the first column (e.g., using the R function order(), then review the data for very large or very small values in that column. Then repeat for each successive column. Another option is to examine the minimum and maximum values of each column using R’s min() and max() functions. For a more automated approach that considers each record as a unit, rather than each column in isolation, clustering techniques (see Chapter 14) could be used to identify clusters of one or a few records that are distant from others. Those records could then be examined.

Missing Values Typically, some records will contain missing values. If the number of records with missing values is small, those records might be omitted. However, if we have a large number of variables, even a small proportion of missing values can affect a lot of records. Even with only 30 variables, if only 5% of the values are missing (spread randomly and independently among cases and variables), almost 80% of the records would have to be omitted from the analysis. (The chance that a given record would escape having a missing value is 0.9530 = 0.215.)

PRELIMINARY STEPS 31

An alternative to omitting records with missing values is to replace the miss- ing value with an imputed value, based on the other values for that variable across all records. For example, if among 30 variables, household income is missing for a particular record, we might substitute the mean household income across all records. Doing so does not, of course, add any information about how house- hold income affects the outcome variable. It merely allows us to proceed with the analysis and not lose the information contained in this record for the other 29 variables. Note that using such a technique will understate the variability in a dataset. However, we can assess variability and the performance of our data mining technique using the validation data, and therefore this need not present a major problem. One option is to replace missing values using fairly simple substitutes (e.g., mean, median). More sophisticated procedures do exist—for example, using linear regression, based on other variables, to fill in the missing values. These methods have been elaborated mainly for analysis of medical and scientific studies, where each patient or subject record comes at great expense. In data mining, where data are typically plentiful, simpler methods usually suf- fice. Table 2.7 shows some R code to illustrate the use of the median to replace

TABLE 2.7 MISSING DATA

code for imputing missing data with median

# To illustrate missing data procedures, we first convert a few entries for # bedrooms to NA's. Then we impute these missing values using the median of the # remaining values. rows.to.missing <- sample(row.names(housing.df), 10) housing.df[rows.to.missing,]$BEDROOMS <- NA summary(housing.df$BEDROOMS) # Now we have 10 NA's and the median of the # remaining values is 3.

# replace the missing values using the median of the remaining values. # use median() with na.rm = TRUE to ignore missing values when computing the median. housing.df[rows.to.missing,]$BEDROOMS <- median(housing.df$BEDROOMS, na.rm = TRUE)

summary(housing.df$BEDROOMS)

Partial Output

> housing.df[rows.to.missing,]$BEDROOMS <- NA > summary(housing.df$BEDROOMS)

Min. 1st Qu. Median Mean 3rd Qu. Max. NA's 1.00 3.00 3.00 3.23 4.00 9.00 10

> housing.df[rows.to.missing,]$BEDROOMS <- median(housing.df$BEDROOMS, na.rm = TRUE) > summary(housing.df$BEDROOMS)

Min. 1st Qu. Median Mean 3rd Qu. Max. 1.00 3.00 3.00 3.23 4.00 9.00

32 OVERVIEW OF THE DATA MINING PROCESS

missing values. Since the data are complete to begin with, the first step is an arti- ficial one of creating some missing records for illustration purposes. The median is used for imputation, rather than the mean, to preserve the integer nature of the counts for bedrooms.

Some datasets contain variables that have a very large number of missing values. In other words, a measurement is missing for a large number of records. In that case, dropping records with missing values will lead to a large loss of data. Imputing the missing values might also be useless, as the imputations are based on a small number of existing records. An alternative is to examine the importance of the predictor. If it is not very crucial, it can be dropped. If it is important, perhaps a proxy variable with fewer missing values can be used instead. When such a predictor is deemed central, the best solution is to invest in obtaining the missing data.

Significant time may be required to deal with missing data, as not all situ- ations are susceptible to automated solutions. In a messy dataset, for example, a “0” might mean two things: (1) the value is missing, or (2) the value is actu- ally zero. In the credit industry, a “0” in the “past due” variable might mean a customer who is fully paid up, or a customer with no credit history at all—two very different situations. Human judgment may be required for individual cases or to determine a special rule to deal with the situation.

Normalizing (Standardizing) and Rescaling Data Some algorithms require that the data be normalized before the algorithm can be implemented effectively. To normalize a variable, we subtract the mean from each value and then divide by the standard deviation. This operation is also sometimes called standardizing. In R, function scale() performs this operation. In effect, we are expressing each value as the “number of standard deviations away from the mean,” also called a z-score.

Normalizing is one way to bring all variables to the same scale. Another popular approach is rescaling each variable to a [0,1] scale. This is done by subtracting the minimum value and then dividing by the range. Subtracting the minimum shifts the variable origin to zero. Dividing by the range shrinks or expands the data to the range [0,1]. In R, rescaling can be done using function rescale() in the scales package.

To consider why normalizing or scaling to [0,1] might be necessary, consider the case of clustering. Clustering typically involves calculating a distance measure that reflects how far each record is from a cluster center or from other records. With multiple variables, different units will be used: days, dollars, counts, and so on. If the dollars are in the thousands and everything else is in the tens, the dollar variable will come to dominate the distance measure. Moreover, changing units from, say, days to hours or months, could alter the outcome completely.

PREDICTIVE POWER AND OVERFITTING 33

Data mining software typically have an option to normalize the data in those algorithms where it may be required. It is an option rather than an automatic feature of such algorithms, because there are situations where we want each variable to contribute to the distance measure in proportion to its original scale.

2.5 Predictive Power and Overfitting

In supervised learning, a key question presents itself: How well will our pre- diction or classification model perform when we apply it to new data? We are particularly interested in comparing the performance of various models so that we can choose the one we think will do the best when it is implemented in prac- tice. A key concept is to make sure that our chosen model generalizes beyond the dataset that we have at hand. To assure generalization, we use the concept of data partitioning and try to avoid overfitting. These two important concepts are described next.

Overfitting

The more variables we include in a model, the greater the risk of overfitting the particular data used for modeling. What is overfitting?

In Table 2.8, we show hypothetical data about advertising expenditures in one time period and sales in a subsequent time period. A scatter plot of the data is shown in Figure 2.2. We could connect up these points with a smooth but complicated function, one that interpolates all these data points perfectly and leaves no error (residuals). This can be seen in Figure 2.3. However, we can see that such a curve is unlikely to be accurate, or even useful, in predicting future sales on the basis of advertising expenditures. For instance, it is hard to believe that increasing expenditures from $400 to $500 will actually decrease revenue.

A basic purpose of building a model is to represent relationships among vari- ables in such a way that this representation will do a good job of predicting future outcome values on the basis of future predictor values. Of course, we want the model to do a good job of describing the data we have, but we are more inter- ested in its performance with future data.

TABLE 2.8

Advertising Sales

239 514 364 789 602 550 644 1386 770 1394 789 1440 911 1354

34 OVERVIEW OF THE DATA MINING PROCESS

FIGURE 2.2 SCATTER PLOT FOR ADVERTISING AND SALES DATA

FIGURE 2.3 OVERFITTING: THIS FUNCTION FITS THE DATA WITH NO ERROR

In the hypothetical advertising example, a simple straight line might do a better job than the complex function in terms of predicting future sales on the basis of advertising. Instead, we devised a complex function that fit the data per- fectly, and in doing so, we overreached. We ended up modeling some variation in the data that is nothing more than chance variation. We mistreated the noise in the data as if it were a signal.

Similarly, we can add predictors to a model to sharpen its performance with the data at hand. Consider a database of 100 individuals, half of whom have contributed to a charitable cause. Information about income, family size, and zip code might do a fair job of predicting whether or not someone is a contributor. If we keep adding additional predictors, we can improve the performance of the model with the data at hand and reduce the misclassification error to a negligible level. However, this low error rate is misleading, because it probably includes spurious effects, which are specific to the 100 individuals, but not beyond that sample.

For example, one of the variables might be height. We have no basis in theory to suppose that tall people might contribute more or less to charity, but if there are several tall people in our sample and they just happened to contribute

PREDICTIVE POWER AND OVERFITTING 35

heavily to charity, our model might include a term for height—the taller you are, the more you will contribute. Of course, when the model is applied to additional data, it is likely that this will not turn out to be a good predictor.

If the dataset is not much larger than the number of predictor variables, it is very likely that a spurious relationship like this will creep into the model. Continuing with our charity example, with a small sample just a few of whom are tall, whatever the contribution level of tall people may be, the algorithm is tempted to attribute it to their being tall. If the dataset is very large relative to the number of predictors, this is less likely to occur. In such a case, each predictor must help predict the outcome for a large number of cases, so the job it does is much less dependent on just a few cases, which might be flukes.

Somewhat surprisingly, even if we know for a fact that a higher-degree curve is the appropriate model, if the model-fitting dataset is not large enough, a lower- degree function (that is not as likely to fit the noise) is likely to perform better in terms of predicting new values. Overfitting can also result from the application of many different models, from which the best performing model is selected.

Creation and Use of Data Partitions

At first glance, we might think it best to choose the model that did the best job of classifying or predicting the outcome variable of interest with the data at hand. However, when we use the same data both to develop the model and to assess its performance, we introduce an “optimism” bias. This is because when we choose the model that works best with the data, this model’s superior performance comes from two sources:

• A superior model

• Chance aspects of the data that happen to match the chosen model better than they match other models

The latter is a particularly serious problem with techniques (such as trees and neural nets) that do not impose linear or other structure on the data, and thus end up overfitting it.

To address the overfitting problem, we simply divide (partition) our data and develop our model using only one of the partitions. After we have a model, we try it out on another partition and see how it performs, which we can measure in several ways. In a classification model, we can count the proportion of held- back records that were misclassified. In a prediction model, we can measure the residuals (prediction errors) between the predicted values and the actual values. This evaluation approach in effect mimics the deployment scenario, where our model is applied to data that it hasn’t “seen.”

We typically deal with two or three partitions: a training set, a validation set, and sometimes an additional test set. Partitioning the data into training,

36 OVERVIEW OF THE DATA MINING PROCESS

validation, and test sets is done either randomly according to predetermined proportions or by specifying which records go into which partition according to some relevant variable (e.g., in time-series forecasting, the data are partitioned according to their chronological order). In most cases, the partitioning should be done randomly to minimize the chance of getting a biased partition. Note the varying nomenclature—the training partition is nearly always called “training” but the names for the other partitions can vary and overlap.

Training Partition The training partition, typically the largest partition, contains the data used to build the various models we are examining. The same training partition is generally used to develop multiple models.

Validation Partition The validation partition (sometimes called the test partition) is used to assess the predictive performance of each model so that you can compare models and choose the best one. In some algorithms (e.g., classi- fication and regression trees, k-nearest neighbors), the validation partition may be used in an automated fashion to tune and improve the model.

Test Partition The test partition (sometimes called the holdout or evalua- tion partition) is used to assess the performance of the chosen model with new data.

Why have both a validation and a test partition? When we use the validation data to assess multiple models and then choose the model that performs best with the validation data, we again encounter another (lesser) facet of the overfitting problem—chance aspects of the validation data that happen to match the cho- sen model better than they match other models. In other words, by using the validation data to choose one of several models, the performance of the chosen model on the validation data will be overly optimistic.

The random features of the validation data that enhance the apparent perfor- mance of the chosen model will probably not be present in new data to which the model is applied. Therefore, we may have overestimated the accuracy of our model. The more models we test, the more likely it is that one of them will be particularly effective in modeling the noise in the validation data. Applying the model to the test data, which it has not seen before, will provide an unbiased estimate of how well the model will perform with new data. Figure 2.4 shows the three data partitions and their use in the data mining process. When we are concerned mainly with finding the best model and less with exactly how well it will do, we might use only training and validation partitions. Table 2.9 shows R code to partition the West Roxbury data into two sets (training and validation) or into three sets (training, validation, and test). This is done by first drawing

PREDICTIVE POWER AND OVERFITTING 37

FIGURE 2.4 THREE DATA PARTITIONS AND THEIR ROLE IN THE DATA MINING PROCESS

a random sample of records into the training set, then assigning the remaining records as validation. In the case of three partitions, the validation records are chosen randomly from the data after excluding the records already sampled into the training set.

Note that with some algorithms, such as nearest-neighbor algorithms, records in the validation and test partitions, and in new data, are compared to records in the training data to find the nearest neighbor(s). As k-nearest neigh- bors is discussed in this book, the use of two partitions is an essential part of the classification or prediction process, not merely a way to improve or assess it. Nonetheless, we can still interpret the error in the validation data in the same way that we would interpret error from any other model.

Cross-Validation When the number of records in our sample is small, data partitioning might not be advisable as each partition will contain too few records for model building and performance evaluation. An alternative to data partitioning is cross-validation, which is especially useful with small samples. Cross-validation is a procedure that starts with partitioning the data into “folds,” or non-overlapping subsamples. Often we choose k = 5 folds, meaning that the data are randomly partitioned into 5 equal parts, where each fold has 20% of the observations. A model is then fit k times. Each time, one of the folds is used as the validation set and the remaining k − 1 folds serve as the training set. The result is that each fold is used once as the validation set, thereby pro- ducing predictions for every observation in the dataset. We can then combine

38 OVERVIEW OF THE DATA MINING PROCESS

TABLE 2.9 DATA PARTITIONING IN R

code for partitioning the West Roxbury data into training, validation (and test) sets

# use set.seed() to get the same partitions when re-running the R code. set.seed(1)

## partitioning into training (60%) and validation (40%) # randomly sample 60% of the row IDs for training; the remaining 40% serve as # validation train.rows <- sample(rownames(housing.df), dim(housing.df)[1]*0.6) # collect all the columns with training row ID into training set: train.data <- housing.df[train.rows, ] # assign row IDs that are not already in the training set, into validation valid.rows <- setdiff(rownames(housing.df), train.rows) valid.data <- housing.df[valid.rows, ]

# alternative code for validation (works only when row names are numeric): # collect all the columns without training row ID into validation set valid.data <- housing.df[-train.rows, ] # does not work in this case

## partitioning into training (50%), validation (30%), test (20%) # randomly sample 50% of the row IDs for training train.rows <- sample(rownames(housing.df), dim(housing.df)[1]*0.5)

# sample 30% of the row IDs into the validation set, drawing only from records # not already in the training set # use setdiff() to find records not already in the training set valid.rows <- sample(setdiff(rownames(housing.df), train.rows),

dim(housing.df)[1]*0.3)

# assign the remaining 20% row IDs serve as test test.rows <- setdiff(rownames(housing.df), union(train.rows, valid.rows))

# create the 3 data frames by collecting all columns from the appropriate rows train.data <- housing.df[train.rows, ] valid.data <- housing.df[valid.rows, ] test.data <- housing.df[test.rows, ]

the model’s predictions on each of the k validation sets in order to evaluate the overall performance of the model. Sometimes cross-validation is built into a data mining algorithm, with the results of the cross-validation used for choosing the algorithm’s parameters (see, e.g., Chapter 9).

2.6 Building a Predictive Model

Let us go through the steps typical to many data mining tasks using a familiar procedure: multiple linear regression. This will help you understand the overall process before we begin tackling new algorithms.

BUILDING A PREDICTIVE MODEL 39

Modeling Process

We now describe in detail the various model stages using the West Roxbury home values example.

1. Determine the purpose. Let’s assume that the purpose of our data mining project is to predict the value of homes in West Roxbury for new records.

2. Obtain the data. We will use the 2014 West Roxbury housing data. The dataset in question is small enough that we do not need to sample from it—we can use it in its entirety.

3. Explore, clean, and preprocess the data. Let’s look first at the description of the variables, also known as the “data dictionary,” to be sure that we understand them all. These descriptions are available on the “descrip- tion” worksheet in the Excel file and in Table 2.2. The variable names and descriptions in this dataset all seem fairly straightforward, but this is not always the case. Often, variable names are cryptic and their descrip- tions may be unclear or missing.

It is useful to pause and think about what the variables mean and whether they should be included in the model. Consider the variable TAX. At first glance, we consider that the tax on a home is usually a function of its assessed value, so there is some circularity in the model— we want to predict a home’s value using TAX as a predictor, yet TAX itself is determined by a home’s value. TAX might be a very good pre- dictor of home value in a numerical sense, but would it be useful if we wanted to apply our model to homes whose assessed value might not be known? For this reason, we will exclude TAX from the analysis.

It is also useful to check for outliers that might be errors. For exam- ple, suppose that the column FLOORS (number of floors) looked like the one in Table 2.10, after sorting the data in descending order based on floors. We can tell right away that the 15 is in error—it is unlikely that a home has 15 floors. Since all other values are between 1 and 2 the decimal was probably misplaced and the value should be 1.5.

Lastly, we create dummy variables for categorical variables. Here we have one categorical variable: REMODEL, which has three categories.

TABLE 2.10 OUTLIER IN WEST ROXBURY DATA

FLOORS ROOMS

15 8 2 10

1.5 6 1 6

40 OVERVIEW OF THE DATA MINING PROCESS

4. Reduce the data dimension. The West Roxbury dataset has been prepared for presentation with fairly low dimension—it has only 13 variables, and the single categorical variable considered has only three categories (and hence adds two dummy variables when used in a linear regression model). If we had many more variables, at this stage we might want to apply a vari- able reduction technique, such as condensing multiple categories into a smaller number, or applying principal components analysis to consolidate multiple similar numerical variables (e.g., LIVING AREA, ROOMS, BEDROOMS, BATH, HALF BATH) into a smaller number of vari- ables.

5. Determine the data mining task. The specific task is to predict the value of TOTAL VALUE using the predictor variables. This is a supervised prediction task. For simplicity, we excluded several additional vari- ables present in the original dataset, which have many categories (BLDG TYPE, ROOF TYPE, and EXT FIN). We therefore use all the numer- ical variables (except TAX) and the dummies created for the remaining categorical variables.

6. Partition the data (for supervised tasks). In this case we divide the data into two partitions: training and validation (see Table 2.9). The training parti- tion is used to build the model, and the validation partition is used to see how well the model does when applied to new data. We need to specify the percent of the data used in each partition. Note: Although not used in our example, a test partition might also be used.

7. Choose the technique. In this case, it is multiple linear regression. Having divided the data into training and validation partitions, we can build a multiple linear regression model with the training data. We want to pre- dict the value of a house in West Roxbury on the basis of all the other predictors (except TAX).

8. Use the algorithm to perform the task. In R, we use the lm() function to predict house value with the training data, then use the same model to predict values for the validation data. Chapter 6 on linear regression goes into more detail. Table 2.11 shows the predicted values for the first few records in the training data along with the actual values and the residuals (prediction errors). Note that the predicted values are often called the fitted values, since they are for the records to which the model was fit. The results for the validation data are shown in Table 2.12. The prediction errors for the training and validation data are compared in Table 2.13.

Prediction error can be aggregated in several ways. Five common mea- sures are shown in Table 2.13. The first is mean error (ME), simply the average of the residuals (errors). In both cases, it is quite small relative to the units of TOTAL VALUE, indicating that, on balance, predictions

BUILDING A PREDICTIVE MODEL 41

TABLE 2.11 PREDICTIONS (FITTED VALUES) FOR A SAMPLE OF TRAINING DATA

code for fitting a regression model to training data (West Roxbury)

reg <- lm(TOTAL_VALUE ~ ., data = housing.df, subset = train.rows) tr.res <- data.frame(train.data$TOTAL_VALUE, reg$fitted.values, reg$residuals) head(tr.res)

Partial Output

> head(tr.res) train.data.TOTAL_VALUE reg.fitted.values reg.residuals

3651 371.6 371.5818 0.018235205 359 299.4 299.4014 -0.001431463 1195 294.5 294.4762 0.023835688 1024 249.4 249.4029 -0.002874472 3984 505.5 505.5246 -0.024612237 2227 410.5 410.5323 -0.032339156

average about right—our predictions are “unbiased.” Of course, this sim- ply means that the positive and negative errors balance out. It tells us nothing about how large these errors are.

TheRMS error (RMSE) (root-mean-squared error) is more informative of the error magnitude: it takes the square root of the average squared error, so it gives an idea of the typical error (whether positive or negative) in the same scale as that used for the original outcome variable. As we might expect, the RMS error for the validation data (161.5 thousand

TABLE 2.12 PREDICTIONS FOR A SAMPLE OF VALIDATION DATA

code for applying the regression model to predict validation set (West Roxbury)

pred <- predict(reg, newdata = valid.data) vl.res <- data.frame(valid.data$TOTAL_VALUE, pred, residuals =

valid.data$TOTAL_VALUE - pred) head(vl.res)

Partial Output

> head(vl.res) valid.data.TOTAL_VALUE pred residuals

1 344.2 344.2388 -0.038842075 14 575.0 575.0025 -0.002509638 16 298.2 298.2107 -0.010697442 17 313.1 313.0764 0.023567596 18 344.9 344.8719 0.028111825 19 330.7 330.7222 -0.022234883

42 OVERVIEW OF THE DATA MINING PROCESS

TABLE 2.13 PREDICTION ERROR METRICS FOR TRAINING AND VALIDATION DATA (ERROR FIGURES ARE IN THOUSANDS OF $)

code for computing model evaluation metrics

library(forecast) # compute accuracy on training set accuracy(reg$fitted.values, train.data$TOTAL_VALUE)

# compute accuracy on prediction set pred <- predict(reg, newdata = valid.data) accuracy(pred, valid.data$TOTAL_VALUE)

Partial Output

> accuracy(reg$fitted.values, train.data$TOTAL_VALUE) ME RMSE MAE MPE MAPE

Test set 1.388101e-16 0.02268016 0.01956465 5.193036e-06 0.00528389

> accuracy(pred, valid.data$TOTAL_VALUE) ME RMSE MAE MPE MAPE

Test set 90.86934 161.5043 118.6455 15.14207 24.45668

dollars), which the model is seeing for the first time in making these predictions, is larger than for the training data (≈ 0 thousand dollars), which were used in training the model. The other measures are discussed in Chapter 5.

9. Interpret the results. At this stage, we would typically try other prediction algorithms (e.g., regression trees) and see how they do error-wise. We might also try different “settings” on the various models (e.g., we could use the best subsets option in multiple linear regression to choose a reduced set of variables that might perform better with the validation data). After choosing the best model—typically, the model with the lowest error on the validation data while also recognizing that “simpler is better”—we use that model to predict the output variable in fresh data. These steps are covered in more detail in the analysis of cases.

10. Deploy the model. After the best model is chosen, it is applied to new data to predict TOTAL VALUE for homes where this value is unknown. This was, of course, the original purpose. Predicting the output value for new records is called scoring. For predictive tasks, scoring produces predicted numerical values. For classification tasks, scoring produces classes and/or propensities. Table 2.14 shows an example of a data frame with three homes to be scored using our linear regression model. Note that all the required predictor columns are present, and the output column is absent.

USING R FOR DATA MINING ON A LOCAL MACHINE 43

TABLE 2.14 DATA FRAME WITH THREE RECORDS TO BE SCORED

# new.data can be read from a csv file, or defined directly in R.

> new.data TAX LOT.SQFT YR.BUILT GROSS.AREA LIVING.AREA FLOORS ROOMS BEDROOMS

100 3818 4200 1960 2670 1710 2.0 10 4 101 3791 6444 1940 2886 1474 1.5 6 3 102 4275 5035 1925 3264 1523 1.0 6 2

FULL.BATH HALF.BATH KITCHEN FIREPLACE REMODEL 100 1 1 1 1 None 101 1 1 1 1 None 102 1 0 1 0 Recent

> pred <- predict(reg, newdata = new.data) > pred

100 101 102 303.5358 301.3919 339.8642

2.7 Using R for Data Mining on a Local Machine

An important aspect of the data mining process is that the heavy-duty analysis does not necessarily require a huge number of records. The dataset to be analyzed may have millions of records, of course, but in applying multiple linear regression or applying a classification tree, the use of a sample of 20,000 is likely to yield as accurate an answer as that obtained when using the entire dataset. The principle involved is the same as the principle behind polling: If sampled judiciously, 2000 voters can give an estimate of the entire population’s opinion within one or two percentage points. (See “How Many Variables and How Much Data” in Section 2.4 for further discussion.)

Therefore, in most cases, the number of records required in each partition (training, validation, and test) can be accommodated within the memory limit allowed in R (to check and increase memory limit in R use function mem- ory.limit()).

When we apply Big Data analytics in R, it might be useful to remove unused objects (function rm()) and call the garbage collection (function gc()) afterwards.

2.8 Automating Data Mining Solutions

In most supervised data mining applications, the goal is not a static, one-time analysis of a particular dataset. Rather, we want to develop a model that can be used on an ongoing basis to predict or classify new records. Our initial analysis will be in prototype mode, while we explore and define the problem and test different models. We will follow all the steps outlined earlier in this chapter.

44 OVERVIEW OF THE DATA MINING PROCESS

At the end of that process, we will typically want our chosen model to be deployed in automated fashion. For example, the US Internal Revenue Service (IRS) receives several hundred million tax returns per year—it does not want to have to pull each tax return out into an Excel sheet or other environment separate from its main database to determine the predicted probability that the return is fraudulent. Rather, it would prefer that determination to be made as part of the normal tax filing environment and process. Music streaming services, such as Pandora or Spotify, need to determine “recommendations” for next songs quickly for each of millions of users; there is no time to extract the data for manual analysis.

In practice, this is done by building the chosen algorithm into the com- putational setting in which the rest of the process lies. A tax return is entered directly into the IRS system by a tax preparer, a predictive algorithm is imme- diately applied to the new data in the IRS system, and a predicted classification is decided by the algorithm. Business rules would then determine what happens with that classification. In the IRS case, the rule might be “if no predicted fraud, continue routine processing; if fraud is predicted, alert an examiner for possible audit.”

This flow of the tax return from data entry, into the IRS system, through a predictive algorithm, then back out to a human user is an example of a “data pipeline.” The different components of the system communicate with one another via Application Programming Interfaces (APIs) that establish locally valid rules for transmitting data and associated communications. An API for a data mining algorithm would establish the required elements for a predictive algorithm to work—the exact predictor variables, their order, data formats, etc. It would also establish the requirements for communicating the results of the algorithm. Algorithms to be used in an automated data pipeline will need to be compliant with the rules of the APIs where they operate.

Finally, once the computational environment is set and functioning, the data miner’s work is not done. The environment in which a model operates is typ- ically dynamic, and predictive models often have a short shelf life—one leading consultant finds they rarely continue to function effectively for more than a year. So, even in a fully deployed state, models must be periodically checked and re- evaluated. Once performance flags, it is time to return to prototype mode and see if a new model can be developed.

In this book, our focus will be on the prototyping phase—all the steps that go into properly defining the model and developing and selecting a model. You should be aware, though, that most of the actual work of implementing a data mining model lies in the automated deployment phase. Much of this work is not in the analytic domain; rather, it lies in the domains of databases and computer science, to assure that detailed nuts and bolts of an automated dataflow all work properly.

AUTOMATING DATA MINING SOLUTIONS 45

D A T A M I N I N G S O F T W A R E : T H E S T A T E O F T H E M A R K E T

by Herb Edelstein∗

The data mining market has changed in some important ways since the last edition of this book. The most significant trends have been the increasing volume of information available and the growing use of the cloud for data storage and analytics. Data mining and analysis have evolved to meet these new demands.

The term “Big Data” reflects the surge in the amount and types of data col- lected. There is still an enormous amount of transactional data, data warehous- ing data, scientific data, and clickstream data. However, adding to the massive storage requirements of traditional data is the influx of information from unstruc- tured sources (e.g., customer service calls and images), social media, and more recently the Internet of Things, which produces a flood of sensor data. The num- ber of organizations collecting such data has greatly increased as virtually every business is expanding in these areas.

Rapid technological change has been an important factor. The price of data storage has dropped precipitously. At this writing, hard disk storage has fallen to about $50 per terabyte and solid state drives are about $200 per terabyte. Concomitant with the decrease in storage costs has been a dramatic increase in bandwidth at ever lower costs. This has enabled the spread of cloud-based computing. Cloud-based computing refers to using remote platforms for storage, data management, and now analysis. Because scaling up the hardware and soft- ware infrastructure for Big Data is so complex, many organizations are entrusting their data to outside vendors. The largest cloud players (Amazon, Google, IBM, and Microsoft) are each reporting annual revenues in excess of US$5 billion, according to Forbes magazine.

Managing this much data is a challenging task. While traditional relational DBMSs—such as Oracle, Microsofts SQL Server, IBMs DB2, and SAPs Adaptive Server Enterprise (formerly Sybase)—are still among the leading data manage- ment tools, open source DBMSs, such as Oracles MySQL are becoming increasingly popular. In addition, nonrelational data storage is making headway in storing extremely large amounts of data. For example, Hadoop, an open source tool for managing large distributed database architectures, has become an important player in the cloud database space. However, Hadoop is a tool for the appli- cation developer community rather than end users. Consequently, many of the data mining analytics vendors such as SAS have built interfaces to Hadoop.

All the major database management system vendors offer data mining capa- bilities, usually integrated into their DBMS. Leading products include Microsoft SQL Server Analysis Services, Oracle Data Mining, and Teradata Warehouse Miner. The target user for embedded data mining is a database professional. Not surpris- ingly, these products take advantage of database functionality, including using the DBMS to transform variables, storing models in the database, and extending the data access language to include model-building and scoring the database. A few products also supply a separate graphical interface for building data min- ing models. Where the DBMS has parallel processing capabilities, embedded data mining tools will generally take advantage of it, resulting in greater performance. As with the data mining suites described below, these tools offer an assortment

46 OVERVIEW OF THE DATA MINING PROCESS

of algorithms. Not only does IBM have embedded analytics in DB2, but follow- ing its acquisition of SPSS, IBM has incorporated Clementine and SPSS into IBM Modeler.

There are still a large number of stand-alone data mining tools based on a single algorithm or on a collection of algorithms called a suite. Target users include both statisticians and business intelligence analysts. The leading suites include SAS Enterprise Miner, SAS JMP, IBM Modeler, Salford Systems SPM, Statis- tica, XLMiner, and RapidMiner. Suites are characterized by providing a wide range of functionality, frequently accessed via a graphical user interface designed to enhance model-building productivity. A popular approach for many of these GUIs is to provide a workflow interface in which the data mining steps and analysis are linked together.

Many suites have outstanding visualization tools and links to statistical packages that extend the range of tasks they can perform. They provide interac- tive data transformation tools as well as a procedural scripting language for more complex data transformations. The suite vendors are working to link their tools more closely to underlying DBMSs; for example, data transformations might be handled by the DBMS. Data mining models can be exported to be incorporated into the DBMS through generating SQL, procedural language code (e.g., C++ or Java), or a standardized data mining model language called Predictive Model Markup Language (PMML).

In contrast to the general-purpose suites, application-specific tools are intended for particular analytic applications such as credit scoring, customer retention, and product marketing. Their focus may be further sharpened to address the needs of specialized markets such as mortgage lending or finan- cial services. The target user is an analyst with expertise in the application domain. Therefore the interfaces, the algorithms, and even the terminology are customized for that particular industry, application, or customer. While less flex- ible than general-purpose tools, they offer the advantage of already incorporating domain knowledge into the product design, and can provide very good solutions with less effort. Data mining companies including SAS, IBM, and RapidMiner offer vertical market tools, as do industry specialists such as Fair Isaac. Other companies, such as Domo, are focusing on creating dashboards with analytics and visualizations for business intelligence.

Another technological shift has occurred with the spread of open source model building tools and open core tools. A somewhat simplified view of open source software is that the source code for the tool is available at no charge to the community of users and can be modified or enhanced by them. These enhancements are submitted to the originator or copyright holder, who can add them to the base package. Open core is a more recent approach in which a core set of functionality remains open and free, but there are proprietary extensions that are not free.

The most important open source statistical analysis software is R. R is descended from a Bell Labs program called S, which was commercialized as S+. Many data mining algorithms have been added to R, along with a plethora of statistics, data management tools, and visualization tools. Because it is essen- tially a programming language, R has enormous flexibility but a steeper learning

AUTOMATING DATA MINING SOLUTIONS 47

curve than many of the GUI-based tools. Although there are some GUIs for R, the overwhelming majority of use is through programming.

Some vendors, as well as the open source community, are adding statisti- cal and data mining tools to Python, a popular programming language that is generally easier to use than C++ or Java, and faster than R.

As mentioned above, the cloud-computing vendors have moved into the data mining/predictive analytics business by offering AaaS (Analytics as a Service) and pricing their products on a transaction basis. These products are oriented more toward application developers than business intelligence analysts. A big part of the attraction of mining data in the cloud is the ability to store and man- age enormous amounts of data without requiring the expense and complexity of building an in-house capability. This can also enable a more rapid implementa- tion of large distributed multi-user applications. Cloud based data can be used with non-cloud-based analytics if the vendors analytics do not meet the users needs.

Amazon has added Amazon Machine Learning to its Amazon Web Services (AWS), taking advantage of predictive modeling tools developed for Amazons internal use. AWS supports both relational databases and Hadoop data manage- ment. Models cannot be exported, because they are intended to be applied to data stored on the Amazon cloud.

Google is very active in cloud analytics with its BigQuery and Prediction API. BigQuery allows the use of Google infrastructure to access large amounts of data using a SQL-like interface. The Prediction API can be accessed from a variety of languages including R and Python. It uses a variety of machine learning algorithms and automatically selects the best results. Unfortunately, this is not a transparent process. Furthermore, as with Amazon, models cannot be exported.

Microsoft is an active player in cloud analytics with its Azure Machine Learn- ing Studio and Stream Analytics. Azure works with Hadoop clusters as well as with traditional relational databases. Azure ML offers a broad range of algorithms such as boosted trees and support vector machines as well as supporting R scripts and Python. Azure ML also supports a workflow interface making it more suit- able for the nonprogrammer data scientist. The real-time analytics component is designed to allow streaming data from a variety of sources to be analyzed on the fly. XLMiner’s cloud version is based on Microsoft Azure. Microsoft also acquired Revolution Analytics, a major player in the R analytics business, with a view to integrating Revolution’s “R Enterprise” with SQL Server and Azure ML. R Enterprise includes extensions to R that eliminate memory limitations and take advantage of parallel processing.

One drawback of the cloud-based analytics tools is a relative lack of trans- parency and user control over the algorithms and their parameters. In some cases, the service will simply select a single model that is a black box to the user. Another drawback is that for the most part cloud-based tools are aimed at more sophisticated data scientists who are systems savvy.

Data science is playing a central role in enabling many organizations to optimize everything from production to marketing. New storage options and analytical tools promise even greater capabilities. The key is to select technology that’s appropriate for an organization’s unique goals and constraints. As always, human judgment is the most important component of a data mining solution.

48 OVERVIEW OF THE DATA MINING PROCESS

This book’s focus is on a comprehensive understanding of the different tech- niques and algorithms used in data mining, and less on the data management requirements of real-time deployment of data mining models. R makes it ideal for this purpose, and for exploration, prototyping, and piloting of solutions.

*** ∗Herb Edelstein is president of Two Crows Consulting (www.twocrows.com), a leading data mining consulting firm near Washington, DC. He is an internationally recognized expert in data mining and data warehousing, a widely published author on these topics, and a popular speaker. Copyright © 2015 Herb Edelstein.

PROBLEMS 49

PROBLEMS

2.1 Assuming that data mining techniques are to be used in the following cases, identify whether the task required is supervised or unsupervised learning.

a. Deciding whether to issue a loan to an applicant based on demographic and financial data (with reference to a database of similar data on prior customers).

b. In an online bookstore, making recommendations to customers concerning addi- tional items to buy based on the buying patterns in prior transactions.

c. Identifying a network data packet as dangerous (virus, hacker attack) based on com- parison to other packets whose threat status is known.

d. Identifying segments of similar customers.

e. Predicting whether a company will go bankrupt based on comparing its financial data to those of similar bankrupt and nonbankrupt firms.

f. Estimating the repair time required for an aircraft based on a trouble ticket.

g. Automated sorting of mail by zip code scanning.

h. Printing of custom discount coupons at the conclusion of a grocery store checkout based on what you just bought and what others have bought previously.

2.2 Describe the difference in roles assumed by the validation partition and the test parti- tion.

2.3 Consider the sample from a database of credit applicants in Table 2.15. Comment on the likelihood that it was sampled randomly, and whether it is likely to be a useful sample.

TABLE 2.15 SAMPLE FROM A DATABASE OF CREDIT APPLICATIONS

OBS CHECK DURATION HISTORY NEW USED FURNITURE RADIO EDUC RETRAIN AMOUNT SAVE RESPONSE ACCT CAR CAR TV ACCT

1 0 6 4 0 0 0 1 0 0 1169 4 1 8 1 36 2 0 1 0 0 0 0 6948 0 1

16 0 24 2 0 0 0 1 0 0 1282 1 0 24 1 12 4 0 1 0 0 0 0 1804 1 1 32 0 24 2 0 0 1 0 0 0 4020 0 1 40 1 9 2 0 0 0 1 0 0 458 0 1 48 0 6 2 0 1 0 0 0 0 1352 2 1 56 3 6 1 1 0 0 0 0 0 783 4 1 64 1 48 0 0 0 0 0 0 1 14421 0 0 72 3 7 4 0 0 0 1 0 0 730 4 1 80 1 30 2 0 0 1 0 0 0 3832 0 1 88 1 36 2 0 0 0 0 1 0 12612 1 0 96 1 54 0 0 0 0 0 0 1 15945 0 0

104 1 9 4 0 0 1 0 0 0 1919 0 1 112 2 15 2 0 0 0 0 1 0 392 0 1

2.4 Consider the sample from a bank database shown in Table 2.16; it was selected ran- domly from a larger database to be the training set. Personal Loan indicates whether a solicitation for a personal loan was accepted and is the response variable. A campaign is planned for a similar solicitation in the future and the bank is looking for a model that will identify likely responders. Examine the data carefully and indicate what your next step would be.

50 OVERVIEW OF THE DATA MINING PROCESS

TABLE 2.16 SAMPLE FROM A BANK DATABASE

OBS AGE EXPERIENCE INCOME ZIP FAMILY CC EDUC MORTGAGE PERSONAL SECURITIES CODE AVG LOAN ACCT

1 25 1 49 91107 4 1.6 1 0 0 1 4 35 9 100 94112 1 2.7 2 0 0 0 5 35 8 45 91330 4 1 2 0 0 0 9 35 10 81 90089 3 0.6 2 104 0 0

10 34 9 180 93023 1 8.9 3 0 1 0 12 29 5 45 90277 3 0.1 2 0 0 0 17 38 14 130 95010 4 4.7 3 134 1 0 18 42 18 81 94305 4 2.4 1 0 0 0 21 56 31 25 94015 4 0.9 2 111 0 0 26 43 19 29 94305 3 0.5 1 97 0 0 29 56 30 48 94539 1 2.2 3 0 0 0 30 38 13 119 94104 1 3.3 2 0 1 0 35 31 5 50 94035 4 1.8 3 0 0 0 36 48 24 81 92647 3 0.7 1 0 0 0 37 59 35 121 94720 1 2.9 1 0 0 0 38 51 25 71 95814 1 1.4 3 198 0 0 39 42 18 141 94114 3 5 3 0 1 1 41 57 32 84 92672 3 1.6 3 0 0 1

2.5 Using the concept of overfitting, explain why when a model is fit to training data, zero error with those data is not necessarily good.

2.6 In fitting a model to classify prospects as purchasers or nonpurchasers, a certain com- pany drew the training data from internal data that include demographic and purchase information. Future data to be classified will be lists purchased from other sources, with demographic (but not purchase) data included. It was found that “refund issued” was a useful predictor in the training data. Why is this not an appropriate variable to include in the model?

2.7 A dataset has 1000 records and 50 variables with 5% of the values missing, spread randomly throughout the records and variables. An analyst decides to remove records with missing values. About how many records would you expect to be removed?

2.8 Normalize the data in Table 2.17, showing calculations.

TABLE 2.17

Age Income ($)

25 49,000 56 156,000 65 99,000 32 192,000 41 39,000 49 57,000

2.9 Statistical distance between records can be measured in several ways. Consider Euclidean distance, measured as the square root of the sum of the squared differences. For the first two records in Table 2.17, it is√

(25− 56)2 + (49, 000− 156, 000)2.

PROBLEMS 51

Can normalizing the data change which two records are farthest from each other in terms of Euclidean distance?

2.10 Two models are applied to a dataset that has been partitioned. Model A is considerably more accurate than model B on the training data, but slightly less accurate than model B on the validation data. Which model are you more likely to consider for final deployment?

2.11 The dataset ToyotaCorolla.csv contains data on used cars on sale during the late summer of 2004 in the Netherlands. It has 1436 records containing details on 38 attributes, including Price, Age, Kilometers, HP, and other specifications.

a. Explore the data using the data visualization capabilities of R.Which of the pairs among the variables seem to be correlated?

b. We plan to analyze the data using various data mining techniques described in future chapters. Prepare the data for use as follows:

i. The dataset has two categorical attributes, Fuel Type andMetallic. Describe how you would convert these to binary variables. Confirm this using R’s functions to transform categorical data into dummies.

ii. Prepare the dataset (as factored into dummies) for data mining techniques of supervised learning by creating partitions in R. Select all the variables and use default values for the random seed and partitioning percentages for training (50%), validation (30%), and test (20%) sets. Describe the roles that these par- titions will play in modeling.

Part II

Data Exploration and Dimension Reduction

CHAPTER 3

Data Visualization

In this chapter, we describe a set of plots that can be used to explore the multidi- mensional nature of a data set. We present basic plots (bar charts, line graphs, and scatter plots), distribution plots (boxplots and histograms), and different enhance- ments that expand the capabilities of these plots to visualize more information. We focus on how the different visualizations and operations can support data mining tasks, from supervised tasks (prediction, classification, and time series forecasting) to unsupervised tasks, and provide a few guidelines on specific visu- alizations to use with each data mining task. We also describe the advantages of interactive visualization over static plots. The chapter concludes with a pre- sentation of specialized plots suitable for data with special structure (hierarchical, network, and geographical).

3.1 Uses of Data Visualization1

The popular saying “a picture is worth a thousand words” refers to the ability to condense diffused verbal information into a compact and quickly understood graphical image. In the case of numbers, data visualization and numerical sum- marization provide us with both a powerful tool to explore data and an effective way to present results.

Where do visualization techniques fit into the data mining process, as described so far? They are primarily used in the preprocessing portion of the data mining process. Visualization supports data cleaning by finding incorrect values (e.g., patients whose age is 999 or −1), missing values, duplicate rows,

1Randall Pruim assisted with the ggplot code in this chapter. This and subsequent sections in this chapter copyright ©2017 Statistics.com and Galit Shmueli. Used by permission.

Data Mining for Business Analytics: Concepts, Techniques, and Applications in R, First Edition. Galit Shmueli, Peter C. Bruce, Inbal Yahav, Nitin R. Patel, and Kenneth C. Lichtendahl, Jr. © 2018 John Wiley & Sons, Inc. Published 2018 by John Wiley & Sons, Inc.

55

56 DATA VISUALIZATION

columns with all the same value, and the like. Visualization techniques are also useful for variable derivation and selection: they can help determine which vari- ables to include in the analysis and which might be redundant. They can also help with determining appropriate bin sizes, should binning of numerical vari- ables be needed (e.g., a numerical outcome variable might need to be converted to a binary variable if a yes/no decision is required). They can also play a role in combining categories as part of the data reduction process. Finally, if the data have yet to be collected and collection is expensive (as with the Pandora project at its outset, see Chapter 7), visualization methods can help determine, using a sample, which variables and metrics are useful.

In this chapter, we focus on the use of graphical presentations for the purpose of data exploration, particularly with relation to predictive analytics. Although our focus is not on visualization for the purpose of data reporting, this chapter offers ideas as to the effectiveness of various graphical displays for the purpose of data presentation. These offer a wealth of useful presentations beyond tabular sum- maries and basic bar charts, which are currently the most popular form of data presentation in the business environment. For an excellent discussion of using graphs to report business data, see Few (2004). In terms of reporting data min- ing results graphically, we describe common graphical displays elsewhere in the book, some of which are technique-specific [e.g., dendrograms for hierarchical clustering (Chapter 15), network charts for social network analysis (Chapter 19), and tree charts for classification and regression trees (Chapter 9)] while others are more general [e.g., receiver operating characteristic (ROC) curves and lift charts for classification (Chapter 5) and profile plots and heatmaps for clustering (Chapter 15)].

Note: The term “graph” can have two meanings in statistics. It can refer, particularly in popular usage, to any of a number of figures to represent data (e.g., line chart, bar plot, histogram, etc.). In a more technical use, it refers to the data structure and visualization in networks (see Chapter 19). Using the term “plot” for the visualizations we explore in this chapter avoids this confusion.

Data exploration is a mandatory initial step whether or not more formal analysis follows. Graphical exploration can support free-form exploration for the purpose of understanding the data structure, cleaning the data (e.g., identifying unexpected gaps or “illegal” values), identifying outliers, discovering initial pat- terns (e.g., correlations among variables and surprising clusters), and generating interesting questions. Graphical exploration can also be more focused, geared toward specific questions of interest. In the data mining context, a combination is needed: free-form exploration performed with the purpose of supporting a specific goal.

Graphical exploration can range from generating very basic plots to using operations such as filtering and zooming interactively to explore a set of interconnected visualizations that include advanced features such as color and

DATA EXAMPLES 57

multiple-panels. This chapter is not meant to be an exhaustive guidebook on visualization techniques, but instead discusses main principles and features that support data exploration in a data mining context. We start by describing vary- ing levels of sophistication in terms of visualization, and show the advantages of different features and operations. Our discussion is from the perspective of how visualization supports the subsequent data mining goal. In particular, we distin- guish between supervised and unsupervised learning; within supervised learning, we also further distinguish between classification (categorical outcome variable) and prediction (numerical outcome variable).

Base R or ggplot?

The ggplot package by Hadley Wickham has become the most popular dedi- cated graphics package in R for creating presentation-quality visualization in a wide variety of contexts. The “gg” in ggplot refers to “Grammar of Graphics,” a term coined by Leland Wilkinson to define a system of plotting theory and nomenclature. Learning ggplot effectively means becoming familiar with this philosophy and technical language of plotting. This brings with it flexibility and power. It also entails a non-trivial learning curve. If you are likely to be using data visualizations on a regular basis in communicating with others with high- quality graphics, it is worth the time and effort to get up to speed on ggplot. If your uses of visualization are likely to be exploratory and informal, perhaps as an initial part of analysis that will not be formally written up, you may decide not to undertake the investment of learning ggplot.

In this chapter, which is primarily an introduction, base R syntax is shown for nearly all the visualizations that we present. In most examples, ggplot syntax is also shown in the R code that follows each Figure.

3.2 Data Examples

To illustrate data visualization, we use two datasets used in additional chapters in the book.

Example 1: Boston Housing Data

The Boston Housing data contain information on census tracts in Boston2 for which several measurements are taken (e.g., crime rate, pupil/teacher ratio). It has 14 variables. A description of each variable is given in Table 3.1 and a sample of the first nine records is shown in Table 3.2. In addition to the original

2The Boston Housing dataset was originally published by Harrison and Rubinfeld in “Hedonic prices and the demand for clean air”, Journal of Environmental Economics & Management, vol. 5, p. 81–102, 1978.

58 DATA VISUALIZATION

TABLE 3.1 DESCRIPTION OF VARIABLES IN BOSTON HOUSING DATASET

CRIM Crime rate

ZN Percentage of residential land zoned for lots over 25,000 ft2

INDUS Percentage of land occupied by nonretail business

CHAS Does tract bound Charles River (= 1 if tract bounds river, = 0 otherwise)

NOX Nitric oxide concentration (parts per 10 million)

RM Average number of rooms per dwelling

AGE Percentage of owner-occupied units built prior to 1940

DIS Weighted distances to five Boston employment centers

RAD Index of accessibility to radial highways

TAX Full-value property tax rate per $10,000

PTRATIO Pupil-to-teacher ratio by town

LSTAT Percentage of lower status of the population

MEDV Median value of owner-occupied homes in $1000s

CAT.MEDV Is median value of owner-occupied homes in tract above $30,000 (CAT.MEDV = 1) or not (CAT.MEDV = 0)

13 variables, the dataset also contains the additional variable CAT.MEDV, which has been created by categorizing median value (MEDV) into two categories: high and low.

We consider three possible tasks:

1. A supervised predictive task, where the outcome variable of interest is the median value of a home in the tract (MEDV).

TABLE 3.2 FIRST NINE RECORDS IN THE BOSTON HOUSING DATA

code for opening the Boston Housing file and viewing the first 9 records

housing.df <- read.csv("BostonHousing.csv") head(housing.df, 9)

Output

> head(housing.df, 9) CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT MEDV CAT.MEDV

1 0.00632 18.0 2.31 0 0.538 6.575 65.2 4.0900 1 296 15.3 4.98 24.0 0 2 0.02731 0.0 7.07 0 0.469 6.421 78.9 4.9671 2 242 17.8 9.14 21.6 0 3 0.02729 0.0 7.07 0 0.469 7.185 61.1 4.9671 2 242 17.8 4.03 34.7 1 4 0.03237 0.0 2.18 0 0.458 6.998 45.8 6.0622 3 222 18.7 2.94 33.4 1 5 0.06905 0.0 2.18 0 0.458 7.147 54.2 6.0622 3 222 18.7 5.33 36.2 1 6 0.02985 0.0 2.18 0 0.458 6.430 58.7 6.0622 3 222 18.7 5.21 28.7 0 7 0.08829 12.5 7.87 0 0.524 6.012 66.6 5.5605 5 311 15.2 12.43 22.9 0 8 0.14455 12.5 7.87 0 0.524 6.172 96.1 5.9505 5 311 15.2 19.15 27.1 0 9 0.21124 12.5 7.87 0 0.524 5.631 100.0 6.0821 5 311 15.2 29.93 16.5 0

BASIC CHARTS: BAR CHARTS, LINE GRAPHS, AND SCATTER PLOTS 59

2. A supervised classification task, where the outcome variable of interest is the binary variable CAT.MEDV that indicates whether the home value is above or below $30,000.

3. An unsupervised task, where the goal is to cluster census tracts.

(MEDV and CAT.MEDV are not used together in any of the three cases).

Example 2: Ridership on Amtrak Trains

Amtrak, a US railway company, routinely collects data on ridership. Here we focus on forecasting future ridership using the series of monthly ridership between January 1991 and March 2004. The data and their source are described in Chapter 16. Hence our task here is (numerical) time series forecasting.

3.3 Basic Charts: Bar Charts, Line Graphs, and Scatter Plots

The three most effective basic plots are bar charts, line graphs, and scatter plots. These plots are easy to create in R and are the plots most commonly used in the current business world, in both data exploration and presentation (unfortunately, pie charts are also popular, although they are usually ineffective visualizations). Basic charts support data exploration by displaying one or two columns of data (variables) at a time. This is useful in the early stages of getting familiar with the data structure, the amount and types of variables, the volume and type of missing values, etc.

The nature of the data mining task and domain knowledge about the data will affect the use of basic charts in terms of the amount of time and effort allo- cated to different variables. In supervised learning, there will be more focus on the outcome variable. In scatter plots, the outcome variable is typically associ- ated with the y-axis. In unsupervised learning (for the purpose of data reduction or clustering), basic plots that convey relationships (such as scatter plots) are pre- ferred.

The top-left panel in Figure 3.1 displays a line chart for the time series of monthly railway passengers on Amtrak. Line graphs are used primarily for show- ing time series. The choice of time frame to plot, as well as the temporal scale, should depend on the horizon of the forecasting task and on the nature of the data.

Bar charts are useful for comparing a single statistic (e.g., average, count, percentage) across groups. The height of the bar (or length in a horizontal display) represents the value of the statistic, and different bars correspond to different groups. Two examples are shown in the bottom panels in Figure 3.1. The left panel shows a bar chart for a numerical variable (MEDV) and the right

60 DATA VISUALIZATION

Year

R id

er sh

ip (

in 0

00 s)

1992 1996 2000 2004

14 00

18 00

22 00

10 20 30

10 20

30 40

50

MDEV

LS TA

T

0 1

CHAS

A vg

. M E

D V

0 5

10 15

20 25

0 1

CHAS

% o

f C AT

.M E

D V

0 5

10 15

20 25

30

FIGURE 3.1 BASIC PLOTS: LINE GRAPH (TOP LEFT), SCATTER PLOT (TOP RIGHT), BAR CHART FOR NUMERICAL VARIABLE (BOTTOM LEFT), AND BAR CHART FOR CATEGORICAL VARIABLE (BOTTOM RIGHT)

code for creating Figure 3.1

## line chart for the Amtrak data Amtrak.df <- read.csv("Amtrak.csv")

# use time series analysis library(forecast) ridership.ts <- ts(Amtrak.df$Ridership, start = c(1991, 1), end = c(2004, 3), freq = 12) plot(ridership.ts, xlab = "Year", ylab = "Ridership (in 000s)", ylim = c(1300, 2300))

## Boston housing data housing.df <- read.csv("BostonHousing.csv")

## scatter plot with axes names plot(housing.df$MEDV ~ housing.df$LSTAT, xlab = "MDEV", ylab = "LSTAT") # alternative plot with ggplot library(ggplot2) ggplot(housing.df) + geom_point(aes(x = LSTAT, y = MEDV), colour = "navy", alpha = 0.7)

## barchart of CHAS vs. mean MEDV # compute mean MEDV per CHAS = (0, 1) data.for.plot <- aggregate(housing.df$MEDV, by = list(housing.df$CHAS), FUN = mean) names(data.for.plot) <- c("CHAS", "MeanMEDV") barplot(data.for.plot$MeanMEDV, names.arg = data.for.plot$CHAS,

xlab = "CHAS", ylab = "Avg. MEDV") # alternative plot with ggplot ggplot(data.for.plot) + geom_bar(aes(x = CHAS, y = MeanMEDV), stat = "identity")

## barchart of CHAS vs. % CAT.MEDV data.for.plot <- aggregate(housing.df$CAT..MEDV, by = list(housing.df$CHAS), FUN = mean) names(data.for.plot) <- c("CHAS", "MeanCATMEDV") barplot(data.for.plot$MeanCATMEDV * 100, names.arg = data.for.plot$CHAS,

xlab = "CHAS", ylab = "% of CAT.MEDV")

BASIC CHARTS: BAR CHARTS, LINE GRAPHS, AND SCATTER PLOTS 61

panel shows a bar chart for a categorical variable (CAT.MEDV). In each, separate bars are used to denote homes in Boston that are near the Charles River vs. those that are not (thereby comparing the two categories of CHAS). The chart with the numerical output MEDV (bottom left) uses the average MEDV on the y- axis. This supports the predictive task: the numerical outcome is on the y-axis and the x-axis is used for a potential categorical predictor.3 (Note that the x-axis on a bar chart must be used only for categorical variables, because the order of bars in a bar chart should be interchangeable.) For the classification task (bottom right), the y-axis indicates the percent of tracts with median value above $30K and the x-axis is a binary variable indicating proximity to the Charles. This graph shows us that the tracts bordering the Charles are much more likely to have median values above $30K.

The top-right panel in Figure 3.1 displays a scatter plot of MEDV vs. LSTAT. This is an important plot in the prediction task. Note that the output MEDV is again on the y-axis (and LSTAT on the x-axis is a potential predic- tor). Because both variables in a basic scatter plot must be numerical, it cannot be used to display the relation between CAT.MEDV and potential predictors for the classification task (but we can enhance it to do so—see Section 3.4). For unsupervised learning, this particular scatter plot helps study the associa- tion between two numerical variables in terms of information overlap as well as identifying clusters of observations.

All three basic plots highlight global information such as the overall level of ridership or MEDV, as well as changes over time (line chart), differences between subgroups (bar chart), and relationships between numerical variables (scatter plot).

Distribution Plots: Boxplots and Histograms

Before moving on to more sophisticated visualizations that enable multidimen- sional investigation, we note two important plots that are usually not considered “basic charts” but are very useful in statistical and data mining contexts. The box- plot and the histogram are two plots that display the entire distribution of a numer- ical variable. Although averages are very popular and useful summary statistics, there is usually much to be gained by looking at additional statistics such as the median and standard deviation of a variable, and even more so by examining the entire distribution. Whereas bar charts can only use a single aggregation, boxplots and histograms display the entire distribution of a numerical variable.

3We refer here to a bar chart with vertical bars. The same principles apply if using a bar chart with horizontal lines, except that the x-axis is now associated with the numerical variable and the y-axis with the categorical variable.

62 DATA VISUALIZATION

Boxplots are also effective for comparing subgroups by generating side-by-side boxplots, or for looking at distributions over time by creating a series of boxplots.

Distribution plots are useful in supervised learning for determining potential data mining methods and variable transformations. For example, skewed numer- ical variables might warrant transformation (e.g., moving to a logarithmic scale) if used in methods that assume normality (e.g., linear regression, discriminant analysis).

A histogram represents the frequencies of all x values with a series of vertical connected bars. For example, in the left panel of Figure 3.2, there are over 150 tracts where the median value (MEDV) is between $20K–$25K.

A boxplot represents the variable being plotted on the y-axis (although the plot can potentially be turned in a 90 degrees angle, so that the boxes are parallel to the x-axis). In the right panel of Figure 3.2 there are two boxplots (called a side-by-side boxplot). The box encloses 50% of the data—for example, in the

Histogram of housing.df$MEDV

MEDV

F re

qu en

cy

10 20 30 40 50

0 50

10 0

15 0

0 1

10 20

30 40

50

CHAS

M E

D V

FIGURE 3.2 DISTRIBUTION CHARTS FOR NUMERICAL VARIABLE MEDV. LEFT: HISTOGRAM, RIGHT: BOXPLOT

code for creating Figure 3.2

## histogram of MEDV hist(housing.df$MEDV, xlab = "MEDV") # alternative plot with ggplot library(ggplot2) ggplot(housing.df) + geom_histogram(aes(x = MEDV), binwidth = 5)

## boxplot of MEDV for different values of CHAS boxplot(housing.df$MEDV ~ housing.df$CHAS, xlab = "CHAS", ylab = "MEDV") # alternative plot with ggplot ggplot(housing.df) + geom_boxplot(aes(x = as.factor(CHAS), y = MEDV)) + xlab("CHAS")

BASIC CHARTS: BAR CHARTS, LINE GRAPHS, AND SCATTER PLOTS 63

right-hand box half of the tracts have median values (MEDV) between $20,000– $33,000. The horizontal line inside the box represents the median (50th per- centile). The top and bottom of the box represent the 75th and 25th percentiles, respectively. Lines extending above and below the box cover the rest of the data range; outliers may be depicted as points or circles. Sometimes the average is marked by a + (or similar) sign. Comparing the average and the median helps in assessing how skewed the data are. Boxplots are often arranged in a series with a different plot for each of the various values of a second variable, shown on the x-axis.

Because histograms and boxplots are geared toward numerical variables, their basic form is useful for prediction tasks. Boxplots can also support unsupervised learning by displaying relationships between a numerical variable (y-axis) and a categorical variable (x-axis). To illustrate these points, look again at Figure 3.2. The left panel shows a histogram of MEDV, revealing a skewed distribution. Transforming the output variable to log(MEDV) might improve results of a linear regression predictor.

The right panel in Figure 3.2 shows side-by-side boxplots comparing the distribution of MEDV for homes that border the Charles River (1) or not (0), similar to Figure 3.1. We see that not only is the average MEDV for river- bounding homes higher than the non-river-bounding homes, the entire distri- bution is higher (median, quartiles, min, and max). We can also see that all river- bounding homes have MEDV above $10 thousand, unlike non-river-bounding homes. This information is useful for identifying the potential importance of this predictor (CHAS), and for choosing data mining methods that can capture the non-overlapping area between the two distributions (e.g., trees).

Boxplots and histograms applied to numerical variables can also provide directions for deriving new variables, for example, they can indicate how to bin a numerical variable (for example, binning a numerical outcome in order to use a naive Bayes classifier, or in the Boston Housing example, choosing the cutoff to convert MEDV to CAT.MEDV).

Finally, side-by-side boxplots are useful in classification tasks for evaluating the potential of numerical predictors. This is done by using the x-axis for the categorical outcome and the y-axis for a numerical predictor. An example is shown in Figure 3.3, where we can see the effects of four numerical predictors on CAT.MEDV. The pairs that are most separated (e.g., PTRATIO and INDUS) indicate potentially useful predictors.

The main weakness of basic charts and distribution plots, in their basic form (that is, using position in relation to the axes to encode values), is that they can only display two variables and therefore cannot reveal high-dimensional infor- mation. Each of the basic charts has two dimensions, where each dimension is dedicated to a single variable. In data mining, the data are usually multivariate by nature, and the analytics are designed to capture and measure multivariate

64 DATA VISUALIZATION

0 1

0. 4

0. 5

0. 6

0. 7

0. 8

CAT.MEDV

N O

X

0 1

10 20

30

CAT.MEDV

LS TA

T

0 1

14 16

18 20

22

CAT.MEDV

P T

R AT

IO

0 1

0 5

10 15

20 25

CAT.MEDV

IN D

U S

FIGURE 3.3 SIDE-BY-SIDE BOXPLOTS FOR EXPLORING THE CAT.MEDV OUTPUT VARIABLE BY DIFFERENT NUMERICAL PREDICTORS. IN A SIDE-BY-SIDE BOXPLOT, ONE AXIS IS USED FOR A CATEGORICAL VARIABLE, AND THE OTHER FOR A NUMERICAL VARIABLE. PLOTTING A CATEGORICAL OUTCOME VARIABLE AND A NUMERICAL PREDICTOR COMPARES THE PREDICTOR’S DISTRIBUTION ACROSS THE OUTCOME CATEGORIES. PLOTTING A NUMERICAL OUTCOME VARIABLE AND A CATEGORICAL PREDICTOR DISPLAYS THE DISTRIBUTION OF THE OUTCOME VARIABLE ACROSS DIFFERENT LEVELS OF THE PREDICTOR

code for creating Figure 3.3

## side-by-side boxplots # use par() to split the plots into panels. par(mfcol = c(1, 4)) boxplot(housing.df$NOX ~ housing.df$CAT..MEDV, xlab = "CAT.MEDV", ylab = "NOX") boxplot(housing.df$LSTAT ~ housing.df$CAT..MEDV, xlab = "CAT.MEDV", ylab = "LSTAT") boxplot(housing.df$PTRATIO ~ housing.df$CAT..MEDV, xlab = "CAT.MEDV", ylab = "PTRATIO") boxplot(housing.df$INDUS ~ housing.df$CAT..MEDV, xlab = "CAT.MEDV", ylab = "INDUS")

information. Visual exploration should therefore also incorporate this important aspect. In the next section, we describe how to extend basic charts (and distribu- tion plots) to multidimensional data visualization by adding features, employing manipulations, and incorporating interactivity. We then present several special- ized charts that are geared toward displaying special data structures (Section 3.5).

Heatmaps: Visualizing Correlations and Missing Values

A heatmap is a graphical display of numerical data where color is used to denote values. In a data mining context, heatmaps are especially useful for two pur- poses: for visualizing correlation tables and for visualizing missing values in the data. In both cases the information is conveyed in a two-dimensional table. A correlation table for p variables has p rows and p columns. A data table contains p columns (variables) and n rows (observations). If the number of rows is huge, then a subset can be used. In both cases, it is much easier and faster to scan the color-coding rather than the values. Note that heatmaps are useful when examining a large number of values, but they are not a replacement for more precise graphical display, such as bar charts, because color differences cannot be perceived accurately.

BASIC CHARTS: BAR CHARTS, LINE GRAPHS, AND SCATTER PLOTS 65

C R

IM Z N

IN D

U S

C H

A S

N O

X

R M

A G

E

D IS

R A

D

TA X

P T

R AT

IO

LS TA

T

M E

D V

C AT

..M E

D V

CAT..MEDV MEDV LSTAT PTRATIO TAX RAD DIS AGE RM NOX CHAS INDUS ZN CRIM

−0.15 0.37 −0.37 0.11 −0.23 0.64 −0.19 0.12 −0.2 −0.27 −0.44 −0.47 0.79 1

−0.39 0.36 −0.48 0.18 −0.43 0.7 −0.38 0.25 −0.38 −0.47 −0.51 −0.74 1 0.79

0.46 −0.41 0.6 −0.05 0.59 −0.61 0.6 −0.5 0.49 0.54 0.37 1 −0.74 −0.47

0.29 −0.39 0.38 −0.12 0.19 −0.36 0.26 −0.23 0.46 0.46 1 0.37 −0.51 −0.44

0.58 −0.31 0.72 −0.04 0.67 −0.29 0.51 −0.53 0.91 1 0.46 0.54 −0.47 −0.27

0.63 −0.31 0.6 −0.01 0.61 −0.21 0.46 −0.49 1 0.91 0.46 0.49 −0.38 −0.2

−0.38 0.66 −0.71 −0.1 −0.77 0.21 −0.75 1 −0.49 −0.53 −0.23 −0.5 0.25 0.12

0.35 −0.57 0.64 0.09 0.73 −0.24 1 −0.75 0.46 0.51 0.26 0.6 −0.38 −0.19

−0.22 0.31 −0.39 0.09 −0.3 1 −0.24 0.21 −0.21 −0.29 −0.36 −0.61 0.7 0.64

0.42 −0.52 0.76 0.09 1 −0.3 0.73 −0.77 0.61 0.67 0.19 0.59 −0.43 −0.23

−0.06 −0.04 0.06 1 0.09 0.09 0.09 −0.1 −0.01 −0.04 −0.12 −0.05 0.18 0.11

0.41 −0.53 1 0.06 0.76 −0.39 0.64 −0.71 0.6 0.72 0.38 0.6 −0.48 −0.37

−0.2 1 −0.53 −0.04 −0.52 0.31 −0.57 0.66 −0.31 −0.31 −0.39 −0.41 0.36 0.37

1 −0.2 0.41 −0.06 0.42 −0.22 0.35 −0.38 0.63 0.58 0.29 0.46 −0.39 −0.15

FIGURE 3.4 HEATMAP OF A CORRELATION TABLE. DARKER VALUES DENOTE STRONGER CORRELATION

code for creating Figure 3.4

## simple heatmap of correlations (without values) heatmap(cor(housing.df), Rowv = NA, Colv = NA)

## heatmap with values library(gplots) heatmap.2(cor(housing.df), Rowv = FALSE, Colv = FALSE, dendrogram = "none",

cellnote = round(cor(housing.df),2), notecol = "black", key = FALSE, trace = 'none', margins = c(10,10))

# alternative plot with ggplot library(ggplot2) library(reshape) # to generate input for the plot cor.mat <- round(cor(housing.df),2) # rounded correlation matrix melted.cor.mat <- melt(cor.mat) ggplot(melted.cor.mat, aes(x = X1, y = X2, fill = value)) +

geom_tile() + geom_text(aes(x = X1, y = X2, label = value))

An example of a correlation table heatmap is shown in Figure 3.4, showing all the pairwise correlations between 13 variables (MEDV and 12 predictors). Darker shades correspond to stronger (positive or negative) correlation. It is easy to quickly spot the high and low correlations.

In a missing value heatmap, rows correspond to records and columns to variables. We use a binary coding of the original dataset where 1 denotes a missing value and 0 otherwise. This new binary table is then colored such that

66 DATA VISUALIZATION

FIGURE 3.5 HEATMAP OF MISSING VALUES IN A DATASET. BLACK DENOTES MISSING VALUE

code for generating a heatmap of missing values

# replace dataFrame with your data. # is.na() returns a Boolean (TRUE/FALSE) output indicating the location of missing # values. # multiplying the Boolean value by 1 converts the output into binary (0/1). heatmap(1 * is.na(dataFrame), Rowv = NA, Colv = NA)

only missing value cells (with value 1) are colored. Figure 3.5 shows an example of a missing value heatmap for a dataset with over 1000 columns. The data include economic, social, political, and “well-being” information on different countries around the world (each row is a country). The variables were merged from multiple sources, and for each source information was not always available on every country. The missing data heatmap helps visualize the level and amount of “missingness” in the merged data file. Some patterns of “missingness” easily emerge: variables that are missing for nearly all observations, as well as clusters of rows (countries) that are missing many values. Variables with little missingness are also visible. This information can then be used for determining how to handle the missingness (e.g., dropping some variables, dropping some records, imputing, or via other techniques).

MULTIDIMENSIONAL VISUALIZATION 67

3.4 Multidimensional Visualization

Basic plots can convey richer information with features such as color, size, and multiple panels, and by enabling operations such as rescaling, aggregation, and interactivity. These additions allow looking at more than one or two variables at a time. The beauty of these additions is their effectiveness in displaying com- plex information in an easily understandable way. Effective features are based on understanding how visual perception works (see Few (2009) for a discussion). The purpose is to make the information more understandable, not just to rep- resent the data in higher dimensions (such as three-dimensional plots that are usually ineffective visualizations).

Adding Variables: Color, Size, Shape, Multiple Panels, and Animation

In order to include more variables in a plot, we must consider the type of variable to include. To represent additional categorical information, the best way is to use hue, shape, or multiple panels. For additional numerical information, we can use color intensity or size. Temporal information can be added via animation.

Incorporating additional categorical and/or numerical variables into the basic (and distribution) plots means that we can now use all of them for both prediction and classification tasks. For example, we mentioned earlier that a basic scatter plot cannot be used for studying the relationship between a cate- gorical outcome and predictors (in the context of classification). However, a very effective plot for classification is a scatter plot of two numerical predictors color-coded by the categorical outcome variable. An example is shown in the top panel of Figure 3.6, with color denoting CAT.MEDV.

In the context of prediction, color-coding supports the exploration of the conditional relationship between the numerical outcome (on the y-axis) and a numerical predictor. Color-coded scatter plots then help assess the need for creating interaction terms (for example, is the relationship between MEDV and LSTAT different for homes near vs. away from the river?).

Color can also be used to include further categorical variables into a bar chart, as long as the number of categories is small. When the number of cat- egories is large, a better alternative is to use multiple panels. Creating multiple panels (also called “trellising”) is done by splitting the observations according to a categorical variable, and creating a separate plot (of the same type) for each category. An example is shown in the right-hand panel of Figure 3.6, where a bar chart of average MEDV by RAD is broken down into two panels by CHAS. We see that the average MEDV for different highway accessibility levels (RAD) behaves differently for homes near the river (lower panel) compared to homes away from the river (upper panel). This is especially salient for RAD = 1. We

68 DATA VISUALIZATION

10 20 30

0 .4

0 .5

0 .6

0 .7

0 .8

LSTAT

N O

X

CAT.MEDV = 1 CAT.MEDV = 0

1 2 3 4 5 6 7 8 24

CHAS = 0

RAD

Av g.

M ED

V

0 15

1 2 3 4 5 6 7 8 24

CHAS = 1

RAD

Av g.

M ED

V

0 30

FIGURE 3.6 ADDING CATEGORICAL VARIABLES BY COLOR-CODING AND MULTIPLE PANELS. LEFT: SCATTER PLOT OF TWO NUMERICAL PREDICTORS, COLOR-CODED BY THE CATEGORICAL OUTCOME CAT.MEDV. RIGHT: BAR CHART OF MEDV BY TWO CATEGORICAL PREDICTORS (CHAS AND RAD), USING MULTIPLE PANELS FOR CHAS

code for creating Figure 3.6

## color plot par(xpd=TRUE) # allow legend to be displayed outside of plot area plot(housing.df$NOX ~ housing.df$LSTAT, ylab = "NOX", xlab = "LSTAT",

col = ifelse(housing.df$CAT..MEDV == 1, "black", "gray")) # add legend outside of plotting area # In legend() use argument inset = to control the location of the legend relative # to the plot. legend("topleft", inset=c(0, -0.2),

legend = c("CAT.MEDV = 1", "CAT.MEDV = 0"), col = c("black", "gray"), pch = 1, cex = 0.5)

# alternative plot with ggplot library(ggplot2) ggplot(housing.df, aes(y = NOX, x = LSTAT, colour= CAT..MEDV)) +

geom_point(alpha = 0.6)

## panel plots # compute mean MEDV per RAD and CHAS # In aggregate() use argument drop = FALSE to include all combinations # (exiting and missing) of RAD X CHAS. data.for.plot <- aggregate(housing.df$MEDV, by = list(housing.df$RAD, housing.df$CHAS),

FUN = mean, drop = FALSE) names(data.for.plot) <- c("RAD", "CHAS", "meanMEDV") # plot the data par(mfcol = c(2,1)) barplot(height = data.for.plot$meanMEDV[data.for.plot$CHAS == 0],

names.arg = data.for.plot$RAD[data.for.plot$CHAS == 0], xlab = "RAD", ylab = "Avg. MEDV", main = "CHAS = 0")

barplot(height = data.for.plot$meanMEDV[data.for.plot$CHAS == 1], names.arg = data.for.plot$RAD[data.for.plot$CHAS == 1], xlab = "RAD", ylab = "Avg. MEDV", main = "CHAS = 1")

# alternative plot with ggplot ggplot(data.for.plot) +

geom_bar(aes(x = as.factor(RAD), y = `meanMEDV`), stat = "identity") + xlab("RAD") + facet_grid(CHAS ~ .)

MULTIDIMENSIONAL VISUALIZATION 69

also see that there are no near-river homes in RAD levels 2, 6, and 7. Such infor- mation might lead us to create an interaction term between RAD and CHAS, and to consider condensing some of the bins in RAD. All these explorations are useful for prediction and classification.

A special plot that uses scatter plots with multiple panels is the scatter plot matrix. In it, all pairwise scatter plots are shown in a single display. The panels in a matrix scatter plot are organized in a special way, such that each column and each row correspond to a variable, thereby the intersections create all the possi- ble pairwise scatter plots. The scatter plot matrix is useful in unsupervised learn- ing for studying the associations between numerical variables, detecting outliers and identifying clusters. For supervised learning, it can be used for examining pairwise relationships (and their nature) between predictors to support variable transformations and variable selection (see Correlation Analysis in Chapter 4). For prediction, it can also be used to depict the relationship of the outcome with the numerical predictors.

An example of a scatter plot matrix is shown in Figure 3.7, with MEDV and three predictors. Below the diagonal are the scatter plots. Variable name indicates the y-axis variable. For example, the plots in the bottom row all have MEDV on the y-axis (which allows studying the individual outcome–predictor

C R

IM IN

D U

S LS

TA T

M E

D V

CRIM INDUS LSTAT MEDV

0.0

0.1

0.2

0.3 Corr:

0.407

Corr:

0.456

Corr:

−0.388

0

10

20 Corr:

0.604

Corr:

−0.484

0

10

20

30 Corr:

−0.738

10 20 30 40 50

0 25 50 75 0 10 20 0 10 20 30 10 20 30 40 50

FIGURE 3.7 SCATTER PLOT MATRIX FOR MEDV AND THREE NUMERICAL PREDICTORS

code for creating Figure 3.7

## simple plot # use plot() to generate a matrix of 4X4 panels with variable name on the diagonal, # and scatter plots in the remaining panels. plot(housing.df[, c(1, 3, 12, 13)])

# alternative, nicer plot (displayed) library(GGally) ggpairs(housing.df[, c(1, 3, 12, 13)])

70 DATA VISUALIZATION

relations). We can see different types of relationships from the different shapes (e.g., an exponential relationship between MEDV and LSTAT and a highly skewed relationship between CRIM and INDUS), which can indicate needed transformations. Along the diagonal, where just a single variable is involved, the frequency distribution for that variable is displayed. Above the diagonal are the correlation coefficients corresponding to the two variables.

Once hue is used, further categorical variables can be added via shape and multiple panels. However, one must proceed cautiously in adding multiple vari- ables, as the display can become over-cluttered and then visual perception is lost.

Adding a numerical variable via size is useful especially in scatter plots (thereby creating “bubble plots”), because in a scatter plot, points represent indi- vidual observations. In plots that aggregate across observations (e.g., boxplots, histograms, bar charts), size and hue are not normally incorporated.

Finally, adding a temporal dimension to a plot to show how the information changes over time can be achieved via animation. A famous example is Rosling’s animated scatter plots showing how world demographics changed over the years (www.gapminder.org). However, while animations of this type work for “sta- tistical storytelling,” they are not very effective for data exploration.

Manipulations: Rescaling, Aggregation and Hierarchies, Zooming, Filtering

Most of the time spent in data mining projects is spent in preprocessing. Typi- cally, considerable effort is expended getting all the data in a format that can actu- ally be used in the data mining software. Additional time is spent processing the data in ways that improve the performance of the data mining procedures. This preprocessing step in data mining includes variable transformation and deriva- tion of new variables to help models perform more effectively. Transformations include changing the numeric scale of a variable, binning numerical variables, and condensing categories in categorical variables. The following manipula- tions support the preprocessing step as well the choice of adequate data mining methods. They do so by revealing patterns and their nature.

Rescaling Changing the scale in a display can enhance the plot and illu- minate relationships. For example, in Figure 3.8, we see the effect of changing both axes of the scatter plot (top) and the y-axis of a boxplot (bottom) to loga- rithmic (log) scale. Whereas the original plots (left) are hard to understand, the patterns become visible in log scale (right). In the histograms, the nature of the relationship between MEDV and CRIM is hard to determine in the original scale, because too many of the points are “crowded” near the y-axis. The re- scaling removes this crowding and allows a better view of the linear relationship between the two log-scaled variables (indicating a log–log relationship). In the boxplot displaying the crowding toward the x-axis in the original units does not allow us to compare the two box sizes, their locations, lower outliers, and most

MULTIDIMENSIONAL VISUALIZATION 71

0 20 40 60 80

10 20

30 40

50

CRIM

M E

D V

0.01 0.10 1.00 10.00 100.00

5 10

20 50

CRIM

M E

D V

0 1

0 20

40 60

80

CAT.MEDV

C R

IM

0 1

0. 01

1. 00

10 0.

0 0

CAT.MEDV

C R

IM

FIGURE 3.8 RESCALING CAN ENHANCE PLOTS AND REVEAL PATTERNS. LEFT: ORIGINAL SCALE, RIGHT: LOGARITHMIC SCALE

code for creating Figure 3.8

options(scipen=999) # avoid scientific notation

## scatter plot: regular and log scale plot(housing.df$MEDV ~ housing.df$CRIM, xlab = "CRIM", ylab = "MEDV") # to use logarithmic scale set argument log = to either 'x', 'y', or 'xy'. plot(housing.df$MEDV ~ housing.df$CRIM,

xlab = "CRIM", ylab = "MEDV", log = 'xy') # alternative log-scale plot with ggplot library(ggplot2) ggplot(housing.df) + geom_point(aes(x = CRIM, y = MEDV)) +

scale_x_log10(breaks = 10^(-2:2), labels = format(10^(-2:2), scientific = FALSE, drop0trailing = TRUE)) +

scale_y_log10(breaks = c(5, 10, 20, 40))

## boxplot: regular and log scale boxplot(housing.df$CRIM ~ housing.df$CAT..MEDV,

xlab = "CAT.MEDV", ylab = "CRIM") boxplot(housing.df$CRIM ~ housing.df$CAT..MEDV,

xlab = "CAT.MEDV", ylab = "CRIM", log = 'y')

of the distribution information. Rescaling removes the “crowding to the x-axis” effect, thereby allowing a comparison of the two boxplots.

Aggregation and Hierarchies Another useful manipulation of scaling is changing the level of aggregation. For a temporal scale, we can aggregate by

72 DATA VISUALIZATION

different granularity (e.g., monthly, daily, hourly) or even by a “seasonal” factor of interest such as month-of-year or day-of-week. A popular aggregation for time series is a moving average, where the average of neighboring values within a given window size is plotted. Moving average plots enhance visualizing a global trend (see Chapter 16).

Non-temporal variables can be aggregated if some meaningful hierarchy exists: geographical (tracts within a zip code in the Boston Housing exam- ple), organizational (people within departments within units), etc. Figure 3.9 illustrates two types of aggregation for the railway ridership time series. The original monthly series is shown in the top-left panel. Seasonal aggregation (by month-of-year) is shown in the top-right panel, where it is easy to see the peak in ridership in July–August and the dip in January–February. The bottom-right panel shows temporal aggregation, where the series is now displayed in yearly aggregates. This plot reveals the global long-term trend in ridership and the generally increasing trend from 1996 on.

Examining different scales, aggregations, or hierarchies supports both super- vised and unsupervised tasks in that it can reveal patterns and relationships at various levels, and can suggest new sets of variables with which to work.

Zooming and Panning The ability to zoom in and out of certain areas of the data on a plot is important for revealing patterns and outliers. We are often interested in more detail on areas of dense information or of special interest. Panning refers to the operation of moving the zoom window to other areas (popular in mapping applications such as Google Maps). An example of zooming is shown in the bottom-left panel of Figure 3.9, where the ridership series is zoomed in to the first two years of the series.

Zooming and panning support supervised and unsupervised methods by detecting areas of different behavior, which may lead to creating new interac- tion terms, new variables, or even separate models for data subsets. In addi- tion, zooming and panning can help choose between methods that assume global behavior (e.g., regression models) and data-driven methods (e.g., expo- nential smoothing forecasters and k-nearest-neighbors classifiers), and indicate the level of global/local behavior (as manifested by parameters such as k in k- nearest neighbors, the size of a tree, or the smoothing parameters in exponential smoothing).

Filtering Filtering means removing some of the observations from the plot. The purpose of filtering is to focus the attention on certain data while eliminating “noise” created by other data. Filtering supports supervised and unsupervised learning in a similar way to zooming and panning: it assists in identifying different or unusual local behavior.

MULTIDIMENSIONAL VISUALIZATION 73

Year

R id

er sh

ip (

in 0

00 s)

1992 1994 1996 1998 2000 2002 2004

14 00

18 00

22 00

Year

R id

er sh

ip (

in 0

00 s)

1991.0 1991.5 1992.0 1992.5

14 00

18 00

22 00

14 00

18 00

22 00

Month

A ve

ra ge

R id

er sh

ip

Jan Mar May Jul Sep Nov

Year

A ve

ra ge

R id

er sh

ip

1992 1994 1996 1998 2000 2002

14 00

18 00

22 00

FIGURE 3.9 TIME SERIES LINE GRAPHS USING DIFFERENT AGGREGATIONS (RIGHT PANELS), ADDING CURVES (TOP-LEFT PANEL), AND ZOOMING IN (BOTTOM-LEFT PANEL)

code for creating Figure 3.9

library(forecast) Amtrak.df <- read.csv("Amtrak data.csv") ridership.ts <- ts(Amtrak.df$Ridership, start = c(1991, 1), end = c(2004, 3), freq = 12)

## fit curve ridership.lm <- tslm(ridership.ts ~ trend + I(trend^2)) plot(ridership.ts, xlab = "Year", ylab = "Ridership (in 000s)", ylim = c(1300, 2300)) lines(ridership.lm$fitted, lwd = 2) # alternative plot with ggplot library(ggplot2) ggplot(Amtrak.df, aes(y = Ridership, x = Month, group = 12)) +

geom_line() + geom_smooth(formula = y ~ poly(x, 2), method= "lm", colour = "navy", se = FALSE, na.rm = TRUE)

## zoom in, monthly, and annual plots ridership.2yrs <- window(ridership.ts, start = c(1991,1), end = c(1992,12)) plot(ridership.2yrs, xlab = "Year", ylab = "Ridership (in 000s)", ylim = c(1300, 2300)) monthly.ridership.ts <- tapply(ridership.ts, cycle(ridership.ts), mean) plot(monthly.ridership.ts, xlab = "Month", ylab = "Average Ridership",

ylim = c(1300, 2300), type = "l", xaxt = 'n') ## set x labels axis(1, at = c(1:12), labels = c("Jan","Feb","Mar", "Apr","May","Jun",

"Jul","Aug","Sep", "Oct","Nov","Dec"))

annual.ridership.ts <- aggregate(ridership.ts, FUN = mean) plot(annual.ridership.ts, xlab = "Year", ylab = "Average Ridership",

ylim = c(1300, 2300))

74 DATA VISUALIZATION

Reference: Trend Lines and Labels

Trend lines and using in-plot labels also help to detect patterns and outliers. Trend lines serve as a reference, and allow us to more easily assess the shape of a pattern. Although linearity is easy to visually perceive, more elaborate rela- tionships such as exponential and polynomial trends are harder to assess by eye. Trend lines are useful in line graphs as well as in scatter plots. An example is shown in the top-left panel of Figure 3.9, where a polynomial curve is overlaid on the original line graph (see also Chapter 16).

In displays that are not overcrowded, the use of in-plot labels can be useful for better exploration of outliers and clusters. An example is shown in Figure 3.10 (a reproduction of Figure 15.1 with the addition of labels). The figure shows different utilities on a scatter plot that compares fuel cost with total sales. We might be interested in clustering the data, and using clustering algorithms to identify clusters that differ markedly with respect to fuel cost and sales. Fig- ure 3.10, with the labels, helps visualize these clusters and their members (e.g., Nevada and Puget are part of a clear cluster with low fuel costs and high sales). For more on clustering and on this example, see Chapter 15.

Scaling up to Large Datasets

When the number of observations (rows) is large, plots that display each indi- vidual observation (e.g., scatter plots) can become ineffective. Aside from using aggregated charts such as boxplots, some alternatives are:

1. Sampling—drawing a random sample and using it for plotting

2. Reducing marker size

3. Using more transparent marker colors and removing fill

4. Breaking down the data into subsets (e.g., by creating multiple panels)

5. Using aggregation (e.g., bubble plots where size corresponds to number of observations in a certain range)

6. Using jittering (slightly moving each marker by adding a small amount of noise)

An example of the advantage of plotting a sample over the large dataset is shown in Figure 12.2 in Chapter 12, where a scatter plot of 5000 records is plotted alongside a scatter plot of a sample. Those plots were generated in Excel. Figure 3.11 illustrates an improved plot of the full dataset by using smaller markers, using jittering to uncover overlaid points, and more transparent colors. We can see that larger areas of the plot are dominated by the grey class, the black class is mainly on the right, while there is a lot of overlap in the top-right area.

MULTIDIMENSIONAL VISUALIZATION 75

5000 10000 15000 20000

0. 5

1. 0

1. 5

2. 0

Sales

F ue

l C os

t

Ariz ona

Bos ton

Cen tral

Com mon

wea lth

NY

Flor ida

Haw aiian

Idah o

Ken tuck

y

Mad ison

Nev ada

New Eng

land

Nort hern

Okla hom

a

Paci fic

Pug et

San Die

go

Sou ther

n

Texa s

Wisc onsi

n

Unit ed

Virg inia

FIGURE 3.10 SCATTER PLOT WITH LABELED POINTS

code for creating Figure 3.10

utilities.df <- read.csv("Utilities.csv")

plot(utilities.df$Fuel_Cost ~ utilities.df$Sales, xlab = "Sales", ylab = "Fuel Cost", xlim = c(2000, 20000))

text(x = utilities.df$Sales, y = utilities.df$Fuel_Cost, labels = utilities.df$Company, pos = 4, cex = 0.8, srt = 20, offset = 0.2)

# alternative with ggplot library(ggplot2) ggplot(utilities.df, aes(y = Fuel_Cost, x = Sales)) + geom_point() +

geom_text(aes(label = paste(" ", Company)), size = 4, hjust = 0.0, angle = 15) + ylim(0.25, 2.25) + xlim(3000, 18000)

Multivariate Plot: Parallel Coordinates Plot

Another approach toward presenting multidimensional information in a two- dimensional plot is via specialized plots such as the parallel coordinates plot. In this

76 DATA VISUALIZATION

10 20 50 100 200

0. 1

0. 2

0. 5

1. 0

2. 0

5. 0

10 .0

Income

C C

A vg

FIGURE 3.11 SCATTER PLOT OF LARGE DATASET WITH REDUCED MARKER SIZE, JITTERING, AND MORE TRANSPARENT COLORING

code for creating Figure 3.11

# use function alpha() in library scales to add transparent colors library(scales) plot(jitter(universal.df$CCAvg, 1) ~ jitter(universal.df$Income, 1),

col = alpha(ifelse(universal.df$Securities.Account == 0, "gray", "black"), 0.4), pch = 20, log = 'xy', ylim = c(0.1, 10), xlab = "Income", ylab = "CCAvg")

# alternative with ggplot library(ggplot2) ggplot(universal.df) +

geom_jitter(aes(x = Income, y = CCAvg, colour = Securities.Account)) + scale_x_log10(breaks = c(10, 20, 40, 60, 100, 200)) + scale_y_log10(breaks = c(0.1, 0.2, 0.4, 0.6, 1.0, 2.0, 4.0, 6.0))

plot a vertical axis is drawn for each variable. Then each observation is repre- sented by drawing a line that connects its values on the different axes, thereby creating a “multivariate profile.” An example is shown in Figure 3.12 for the Boston Housing data. In this display, separate panels are used for the two values of CAT.MEDV, in order to compare the profiles of homes in the two classes (for a classification task). We see that the more expensive homes (bottom panel) consistently have low CRIM, low LSAT, and high RM compared to cheaper

MULTIDIMENSIONAL VISUALIZATION 77

CAT.MEDV = 0

CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT MEDV

CAT.MEDV = 1

CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT MEDV

FIGURE 3.12 PARALLEL COORDINATES PLOT FOR BOSTON HOUSING DATA. EACH OF THE VARIABLES (SHOWN ON THE HORIZONTAL AXIS) IS SCALED TO 0–100%. PANELS ARE USED TO DISTINGUISH CAT.MEDV (TOP PANEL = HOMES BELOW $30,000)

code for creating Figure 3.12

library(MASS) par(mfcol = c(2,1)) parcoord(housing.df[housing.df$CAT..MEDV == 0, -14], main = "CAT.MEDV = 0") parcoord(housing.df[housing.df$CAT..MEDV == 1, -14], main = "CAT.MEDV = 1")

homes (top panel), which are more mixed on CRIM, and LSAT, and have a medium level of RM. This observation gives indication of useful predictors and suggests possible binning for some numerical predictors.

Parallel coordinates plots are also useful in unsupervised tasks. They can reveal clusters, outliers, and information overlap across variables. A useful manip- ulation is to reorder the columns to better reveal observation clusterings.

Interactive Visualization

Similar to the interactive nature of the data mining process, interactivity is key to enhancing our ability to gain information from graphical visualization. In the words of Stephen Few (Few, 2009), an expert in data visualization:

We can only learn so much when staring at a static visualization such as a printed graph …If we can’t interact with the data …we hit the wall.

By interactive visualization, we mean an interface that supports the following principles:

1. Making changes to a chart is easy, rapid, and reversible.

2. Multiple concurrent charts and tables can be easily combined and dis- played on a single screen.

78 DATA VISUALIZATION

3. A set of visualizations can be linked, so that operations in one display are reflected in the other displays.

Let us consider a few examples where we contrast a static plot generator (e.g., Excel) with an interactive visualization interface.

Histogram rebinning Consider the need to bin a numerical variables and using a histogram for that purpose. A static histogram would require replotting for each new binning choice. If the user generates multiple plots, then the screen becomes cluttered. If the same plot is recreated, then it is hard to compare different binning choices. In contrast, an interactive visualization would provide an easy way to change bin width interactively (see, e.g., the slider below the histogram in Figure 3.13), and then the histogram would automatically and rapidly replot as the user changes the bin width.

Aggregation and Zooming Consider a time series forecasting task, given a long series of data. Temporal aggregation at multiple levels is needed for deter- mining short and long term patterns. Zooming and panning are used to iden- tify unusual periods. A static plotting software requires the user to compute new variables for each temporal aggregation (e.g., aggregate daily data to obtain weekly aggregates). Zooming and panning requires manually changing the min and max values on the axis scale of interest (thereby losing the ability to quickly move between different areas without creating multiple charts). An interactive visualization would provide immediate temporal hierarchies which the user can easily switch between. Zooming would be enabled as a slider near the axis (see, e.g., the sliders on the top-left panel in Figure 3.13), thereby allowing direct manipulation and rapid reaction.

Combining Multiple Linked Plots That Fit in a Single Screen To support a clas- sification task, multiple plots are created of the outcome variable vs. potential categorical and numerical predictors. These can include side-by-side boxplots, color-coded scatter plots, and multipanel bar charts. The user wants to detect possible multidimensional relationships (and identify possible outliers) by select- ing a certain subset of the data (e.g., a single category of some variable) and locating the observations on the other plots. In a static interface, the user would have to manually organize the plots of interest and resize them in order to fit within a single screen. A static interface would usually not support inter-plot linkage, and even if it did, the entire set of plots would have to be regenerated each time a selection is made. In contrast, an interactive visualization would provide an easy way to automatically organize and resize the set of plots to fit within a screen. Linking the set of plots would be easy, and in response to the users selection on one plot, the appropriate selection would be automatically highlighted in the other plots (see example in Figure 3.13).

MULTIDIMENSIONAL VISUALIZATION 79

FIGURE 3.13 MULTIPLE INTER-LINKED PLOTS IN A SINGLE VIEW (USING SPOTFIRE). NOTE THE MARKED OBSERVATION IN THE TOP-LEFT PANEL, WHICH IS ALSO HIGHLIGHTED IN ALL OTHER PLOTS

In earlier sections, we used plots to illustrate the advantages of visualizations, because “a picture is worth a thousand words.” The advantages of an interactive visualization are even harder to convey in words. As Ben Shneiderman, a well- known researcher in information visualization and interfaces, puts it:

A picture is worth a thousand words. An interface is worth a thousand pictures.

The ability to interact with plots, and link them together turns plotting into an analytical tool that supports continuous exploration of the data. Sev- eral commercial visualization tools provide powerful capabilities along these lines; two very popular ones are Spotfire (http://spotfire.tibco.com) and Tableau (www.tableausoftware.com); Figure 3.13 was generated using Spotfire.

Tableau and Spotfire have spent hundreds of millions of dollars on software R&D and review of interactions with customers to hone interfaces that allow analysts to interact with their data via plots smoothly and efficiently. It is difficult to replicate in a programming language like R the sophisticated and highly engi- neered user interface required for rapid progression through different exploratory views of the data. The need is there, however, and the R community is moving

80 DATA VISUALIZATION

to provide the capability to provide interactivity in plots. The widespread use of javascript in web development has led some programmers to provide R wrap- pers for javascript plotting tools such as Highcharts and Plotly. See “Interactive Charts in R” at http://flowingdata.com/2016/10/21/interactive-charts-in-r/. The tool ggvis, from Hadley Wickham at RStudio, also provides interactivity capabilities for R plots. As of the time of this writing, interactivity tools in R were evolving, and R programmers are likely to see more and higher level tools that will allow them to develop custom interactive plots quickly, plots that can be deployed and used by other non-programming analysts in the organization.

3.5 Specialized Visualizations

In this section, we mention a few specialized visualizations that are able to capture data structures beyond the standard time series and cross-sectional structures— special types of relationships that are usually hard to capture with ordinary plots. In particular, we address hierarchical data, network data, and geographical data— three types of data that are becoming increasingly available.

Visualizing Networked Data

Network analysis techniques were spawned by the explosion of social and prod- uct network data. Examples of social networks are networks of sellers and buyers on eBay and networks of users on Facebook. An example of a product network is the network of products on Amazon (linked through the recommendation system). Network data visualization is available in various network-specialized software, and also in general-purpose software.

A network diagram consists of actors and relations between them. “Nodes” are the actors (e.g., users in a social network or products in a product network), and represented by circles. “Edges” are the relations between nodes, and are represented by lines connecting nodes. For example, in a social network such as Facebook, we can construct a list of users (nodes) and all the pairwise relations (edges) between users who are “Friends.” Alternatively, we can define edges as a posting that one user posts on another user’s Facebook page. In this setup, we might have more than a single edge between two nodes. Networks can also have nodes of multiple types. A common structure is networks with two types of nodes. An example of a two-type node network is shown in Figure 3.14, where we see a set of transactions between a network of sellers and buyers on the online auction site www.eBay.com [the data are for auctions selling Swarovski beads, and took place during a period of several months; from Jank and Yahav (2010)]. The black circles represent sellers and the grey circles represent buyers. Circle size represents the number of transactions that the node (seller or buyer) was involved in within this network. Line width represents the number of auctions that the

SPECIALIZED VISUALIZATIONS 81

FIGURE 3.14 NETWORK GRAPH OF EBAY SELLERS (BLACK CIRCLES) AND BUYERS (GREY CIRCLES) OF SWAROVSKI BEADS. CIRCLE SIZE REPRESENTS THE NODE’S NUMBER OF TRANSACTIONS. LINE WIDTH REPRESENTS THE NUMBER OF TRANSACTIONS BETWEEN THAT PAIR OF SELLER–BUYER

code for creating Figure 3.14

library(igraph) ebay.df <- read.csv("eBayNetwork.csv")

# transform node ids to factors ebay.df[,1] <- as.factor(ebay.df[,1]) ebay.df[,2] <- as.factor(ebay.df[,2])

graph.edges <- as.matrix(ebay.df[,1:2]) g <- graph.edgelist(graph.edges, directed = FALSE) isBuyer <- V(g)$name %in% graph.edges[,2]

plot(g, vertex.label = NA, vertex.color = ifelse(isBuyer, "gray", "black"), vertex.size = ifelse(isBuyer, 7, 10))

bidder–seller pair interacted in. We can see that this marketplace is dominated by three or four high-volume sellers. We can also see that many buyers interact with a single seller. The market structures for many individual products could

82 DATA VISUALIZATION

be reviewed quickly in this way. Network providers could use the information, for example, to identify possible partnerships to explore with sellers.

Figure 3.14 was produced using R’s igraph package. Another useful pack- age, especially for social network analysis, is sna. Using these packages, networks can be imported from social network websites such as Twitter and Facebook. The graph’s appearance can be customized and various features are available such as filtering nodes and edges, 3D visualization, altering the graph’s layout, find- ing clusters of related nodes, calculating graph metrics, and performing network analysis (see Chapter 19 for details and examples).

Network graphs can be potentially useful in the context of association rules (see Chapter 14). For example, consider a case of mining a dataset of consumers’ grocery purchases to learn which items are purchased together (“what goes with what”). A network can be constructed with items as nodes and edges connecting items that were purchased together. After a set of rules is generated by the data mining algorithm (which often contains an excessive number of rules, many of which are unimportant), the network graph can help visualize different rules for the purpose of choosing the interesting ones. For example, a popular “beer and diapers” combination would appear in the network graph as a pair of nodes with very high connectivity. An item which is almost always purchased regardless of other items (e.g., milk) would appear as a very large node with high connectivity to all other nodes.

Visualizing Hierarchical Data: Treemaps

We discussed hierarchical data and the exploration of data at different hierarchy levels in the context of plot manipulations. Treemaps are useful visualizations specialized for exploring large data sets that are hierarchically structured (tree- structured). They allow exploration of various dimensions of the data while maintaining the hierarchical nature of the data. An example is shown in Fig- ure 3.15, which displays a large set of auctions from eBay.com,4 hierarchically ordered by item category, sub-category, and brand. The levels in the hierarchy of the treemap are visualized as rectangles containing sub-rectangles. Categorical variables can be included in the display by using hue. Numerical variables can be included via rectangle size and color intensity (ordering of the rectangles is sometimes used to reinforce size). In the example in Figure 3.15, size is used to represent the average closing price (which reflects item value) and color intensity represents the percent of sellers with negative feedback (a negative seller feed- back indicates buyer dissatisfaction in past transactions and is often indicative of fraudulent seller behavior). Consider the task of classifying ongoing auctions in terms of a fraudulent outcome. From the treemap, we see that the highest

4We thank Sharad Borle for sharing this dataset.

SPECIALIZED VISUALIZATIONS 83

negative.feedback 0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014

Staplers

Dewalt_Cordless_Drill

Bausch_and_Laumb_Microscope

Oakley_Sunglasses

Brioni_Tie Zegna_TieCross_PenWaterman_Pen

HP_Inkjet_Color_Printer

Casio_Calculator Sharp_Calculator

Celestron_Telescope

Hair_DryerShaver

Cartier_WristwatchRolex_Wristwatch

Samsonite_Luggage

Rookwood_Vase

Roseville_Vase

Callaway_Golfbag

Callaway_Golfballs

Ping_Golfbag

Titleist_Golfballs

Desktop accessoriesElectric Drills

Microscopes

Sunglasses

Neck tiesPremium Pens

Computer AccessoriesCalculators

Telescopes

Hair Care

Men's electric shavers

Premium wristwatches

Luggage bags

Collectible Pottery Golf

Business & Industrial

Clothing & accessories

Clothing shoes & accessories

Collectibles

ComputersConsumer Electronics

Health & Beauty

Jewelry & watches

Luggage

Pottery & Glass Sports

FIGURE 3.15 TREEMAP SHOWING NEARLY 11,000 EBAY AUCTIONS, ORGANIZED BY ITEM CATEGORY, SUBCATEGORY, AND BRAND. RECTANGLE SIZE REPRESENTS AVERAGE CLOSING PRICE (REFLECTING ITEM VALUE). SHADE REPRESENTS PERCENTAGE OF SELLERS WITH NEGATIVE FEEDBACK (DARKER = HIGHER)

code for creating Figure 3.15

library(treemap) tree.df <- read.csv("EbayTreemap.csv")

# add column for negative feedback tree.df$negative.feedback <- 1* (tree.df$Seller.Feedback < 0)

# draw treemap treemap(tree.df, index = c("Category","Sub.Category", "Brand"),

vSize = "High.Bid", vColor = "negative.feedback", fun.aggregate = "mean", align.labels = list(c("left", "top"), c("right", "bottom"), c("center", "center")), palette = rev(gray.colors(3)), type = "manual", title = "")

proportion of sellers with negative ratings (black) is concentrated in expensive item auctions (Rolex and Cartier wristwatches).

Ideally, treemaps should be explored interactively, zooming to different levels of the hierarchy. One example of an interactive online application of treemaps is currently available at www.drasticdata.nl. One of their treemap examples dis- plays player-level data from the 2014 World Cup, aggregated to team level. The user can choose to explore players and team data.

Visualizing Geographical Data: Map Charts

Many datasets used for data mining now include geographical information. Zip codes are one example of a categorical variable with many categories, where it is not straightforward to create meaningful variables for analysis. Plotting the data on a geographic map can often reveal patterns that are harder to identify

84 DATA VISUALIZATION

otherwise. A map chart uses a geographical map as its background; then color, hue, and other features are used to include categorical or numerical variables. Besides specialized mapping software, maps are now becoming part of general- purpose software, and Google Maps provides APIs (application programming interfaces) that allow organizations to overlay their data on a Google map. While Google Maps is readily available, resulting map charts (such as Figure 3.16) are

0

20

40

60

−150 −125 −100 −75 −50 lon

la t

FIGURE 3.16 MAP CHART OF STUDENTS’ AND INSTRUCTORS’ LOCATIONS ON A GOOGLE MAP (FROM STATISTICS.COM)

code for creating Figure 3.16

library(ggmap) SCstudents <- read.csv("SC-US-students-GPS-data-2016.csv") Map <- get_map("Denver, CO", zoom = 3) ggmap(Map) + geom_point(aes(x = longitude, y = latitude), data = SCstudents,

alpha = 0.4, colour = "red", size = 0.5)

SPECIALIZED VISUALIZATIONS 85

somewhat inferior in terms of effectiveness compared to map charts in dedicated interactive visualization software.

Figure 3.17 shows two world map charts (created with Spotfire), comparing countries’ “well-being” (according to a 2006 Gallup survey) in the top map, to gross domestic product (GDP) in the bottom map. Lighter shade means higher value.

4

6

8

Happiness

5.0e+12

1.0e+13

1.5e+13

GDP2015

FIGURE 3.17 WORLD MAPS COMPARING “WELL-BEING” (TOP) TO GDP (BOTTOM). SHADING BY AVERAGE “GLOBAL WELL-BEING” SCORE (TOP) OR GDP (BOTTOM) OF COUNTRY. LIGHTER CORRESPONDS TO HIGHER SCORE OR LEVEL. DATA FROM VEENHOVEN’S WORLD DATABASE OF HAPPINESS

code for creating Figure 3.17

library(mosaic)

gdp.df <- read.csv("gdp.csv", skip = 4, stringsAsFactors = FALSE) names(gdp.df)[5] <- "GDP2015" happiness.df <- read.csv("Veerhoven.csv")

# gdp map mWorldMap(gdp.df, key = "Country.Name", fill = "GDP2015") + coord_map()

# eell-being map mWorldMap(happiness.df, key = "Nation", fill = "Score") + coord_map() +

scale_fill_continuous(name = "Happiness")

86 DATA VISUALIZATION

3.6 Summary: Major Visualizations and Operations, by Data Mining Goal

Prediction

• Plot outcome on the y-axis of boxplots, bar charts, and scatter plots.

• Study relation of outcome to categorical predictors via side-by-side box- plots, bar charts, and multiple panels.

• Study relation of outcome to numerical predictors via scatter plots.

• Use distribution plots (boxplot, histogram) for determining needed trans- formations of the outcome variable (and/or numerical predictors).

• Examine scatter plots with added color/panels/size to determine the need for interaction terms.

• Use various aggregation levels and zooming to determine areas of the data with different behavior, and to evaluate the level of global vs. local patterns.

Classification

• Study relation of outcome to categorical predictors using bar charts with the outcome on the y-axis.

• Study relation of outcome to pairs of numerical predictors via color-coded scatter plots (color denotes the outcome).

• Study relation of outcome to numerical predictors via side-by-side box- plots: Plot boxplots of a numerical variable by outcome. Create similar displays for each numerical predictor. The most separable boxes indicate potentially useful predictors.

• Use color to represent the outcome variable on a parallel coordinate plot.

• Use distribution plots (boxplot, histogram) for determining needed trans- formations of numerical predictor variables.

• Examine scatter plots with added color/panels/size to determine the need for interaction terms.

• Use various aggregation levels and zooming to determine areas of the data with different behavior, and to evaluate the level of global vs. local patterns.

Time Series Forecasting

• Create line graphs at different temporal aggregations to determine types of patterns.

• Use zooming and panning to examine various shorter periods of the series to determine areas of the data with different behavior.

SUMMARY: MAJOR VISUALIZATIONS AND OPERATIONS, BY DATA MINING GOAL 87

• Use various aggregation levels to identify global and local patterns.

• Identify missing values in the series (that will require handling).

• Overlay trend lines of different types to determine adequate modeling choices.

Unsupervised Learning

• Create scatter plot matrices to identify pairwise relationships and cluster- ing of observations.

• Use heatmaps to examine the correlation table.

• Use various aggregation levels and zooming to determine areas of the data with different behavior.

• Generate a parallel coordinates plot to identify clusters of observations.

88 DATA VISUALIZATION

PROBLEMS

3.1 Shipments of Household Appliances: Line Graphs. The file ApplianceShip- ments.csv contains the series of quarterly shipments (in millions of dollars) of US house- hold appliances between 1985 and 1989.

a. Create a well-formatted time plot of the data using R.

b. Does there appear to be a quarterly pattern? For a closer view of the patterns, zoom in to the range of 3500–5000 on the y-axis.

c. Using R, create one chart with four separate lines, one line for each of Q1, Q2, Q3, and Q4. In R, this can be achieved by generating a data.frame for each quarter Q1, Q2, Q3, Q4, and then plotting them as separate series on the line graph. Zoom in to the range of 3500–5000 on the y-axis. Does there appear to be a difference between quarters?

d. Using R, create a line graph of the series at a yearly aggregated level (i.e., the total shipments in each year).

3.2 Sales of Riding Mowers: Scatter Plots. A company that manufactures riding mowers wants to identify the best sales prospects for an intensive sales campaign. In particular, the manufacturer is interested in classifying households as prospective own- ers or nonowners on the basis of Income (in $1000s) and Lot Size (in 1000 ft2). The marketing expert looked at a random sample of 24 households, given in the file Riding- Mowers.csv.

a. Using R, create a scatter plot of Lot Size vs. Income, color-coded by the outcome variable owner/nonowner. Make sure to obtain a well-formatted plot (create legi- ble labels and a legend, etc.).

3.3 Laptop Sales at a London Computer Chain: Bar Charts and Boxplots. The file LaptopSalesJanuary2008.csv contains data for all sales of laptops at a computer chain in London in January 2008. This is a subset of the full dataset that includes data for the entire year.

a. Create a bar chart, showing the average retail price by store. Which store has the highest average? Which has the lowest?

b. To better compare retail prices across stores, create side-by-side boxplots of retail price by store. Now compare the prices in the two stores from (a). Does there seem to be a difference between their price distributions?

3.4 Laptop Sales at a London Computer Chain: Interactive Visualization. The next exercises are designed for using an interactive visualization tool. The file LaptopSales.txt is a comma-separated file with nearly 300,000 rows. ENBIS (the European Network for Business and Industrial Statistics) provided these data as part of a contest organized in the fall of 2009. Scenario: Imagine that you are a new analyst for a company called Acell (a company selling laptops). You have been provided with data about products and sales. You need to help the company with their business goal of planning a product strategy and pricing policies that will maximize Acell’s projected revenues in 2009. Using an interactive visualization tool, answer the following questions.

a. Price Questions:

i. At what price are the laptops actually selling?

ii. Does price change with time? (Hint: Make sure that the date column is recog- nized as such. The software should then enable different temporal aggregation

PROBLEMS 89

choices, e.g., plotting the data by weekly or monthly aggregates, or even by day of week.)

iii. Are prices consistent across retail outlets?

iv. How does price change with configuration?

b. Location Questions:

i. Where are the stores and customers located?

ii. Which stores are selling the most?

iii. How far would customers travel to buy a laptop?

◦ Hint 1: You should be able to aggregate the data, for example, plot the sum or average of the prices.

◦ Hint 2: Use the coordinated highlighting between multiple visualizations in the same page, for example, select a store in one view to see the matching customers in another visualization.

◦ Hint 3: Explore the use of filters to see differences. Make sure to filter in the zoomed out view. For example, try to use a “store location” slider as an alternative way to dynamically compare store locations. This might be more useful to spot outlier patterns if there were 50 store locations to compare.

iv. Try an alternative way of looking at how far customers traveled. Do this by creating a new data column that computes the distance between customer and store.

c. Revenue Questions:

i. How do the sales volume in each store relate to Acell’s revenues?

ii. How does this relationship depend on the configuration?

d. Configuration Questions:

i. What are the details of each configuration? How does this relate to price?

ii. Do all stores sell all configurations?

CHAPTER 4

Dimension Reduction

In this chapter, we describe the important step of dimension reduction. The dimension of a dataset, which is the number of variables, must be reduced for the data mining algorithms to operate efficiently. This process is part of the pilot/prototype phase of data mining and is done before deploying a model. We present and discuss several dimension reduction approaches: (1) Incorporat- ing domain knowledge to remove or combine categories, (2) using data sum- maries to detect information overlap between variables (and remove or combine redundant variables or categories), (3) using data conversion techniques such as converting categorical variables into numerical variables, and (4) employing automated reduction techniques, such as principal components analysis (PCA), where a new set of variables (which are weighted averages of the original vari- ables) is created. These new variables are uncorrelated and a small subset of them usually contains most of their combined information (hence, we can reduce dimension by using only a subset of the new variables). Finally, we mention data mining methods such as regression models and classification and regression trees, which can be used for removing redundant variables and for combining “similar” categories of categorical variables.

4.1 Introduction

In data mining, one often encounters situations where there is a large number of variables in the database. Even when the initial number of variables is small, this set quickly expands in the data preparation step, where new derived variables are created (e.g., dummies for categorical variables and new forms of existing variables). In such situations, it is likely that subsets of variables are highly cor- related with each other. Including highly correlated variables in a classification

91

92 DIMENSION REDUCTION

or prediction model, or including variables that are unrelated to the outcome of interest, can lead to overfitting, and accuracy and reliability can suffer. A large number of variables also poses computational problems for some supervised as well as unsupervised algorithms (aside from questions of correlation). In model deployment, superfluous variables can increase costs due to the collection and processing of these variables.

4.2 Curse of Dimensionality

The dimensionality of a model is the number of predictors or input variables used by the model. The curse of dimensionality is the affliction caused by adding vari- ables to multivariate data models. As variables are added, the data space becomes increasingly sparse, and classification and prediction models fail because the avail- able data are insufficient to provide a useful model across so many variables. An important consideration is the fact that the difficulties posed by adding a variable increase exponentially with the addition of each variable. One way to think of this intuitively is to consider the location of an object on a chessboard. It has two dimensions and 64 squares or choices. If you expand the chessboard to a cube, you increase the dimensions by 50%—from 2 dimensions to 3 dimen- sions. However, the location options increase by 800%, to 512 (8 × 8 × 8). In statistical distance terms, the proliferation of variables means that nothing is close to anything else anymore—too much noise has been added and patterns and structure are no longer discernible. The problem is particularly acute in Big Data applications, including genomics, where, for example, an analysis might have to deal with values for thousands of different genes. One of the key steps in data mining, therefore, is finding ways to reduce dimensionality with minimal sacrifice of accuracy. In the artificial intelligence literature, dimension reduction is often referred to as factor selection or feature extraction.

4.3 Practical Considerations

Although data mining prefers automated methods over domain knowledge, it is important at the first step of data exploration to make sure that the variables mea- sured are reasonable for the task at hand. The integration of expert knowledge through a discussion with the data provider (or user) will probably lead to better results. Practical considerations include: Which variables are most important for the task at hand, and which are most likely to be useless? Which variables are likely to contain much error? Which variables will be available for mea- surement (and what will it cost to measure them) in the future if the analysis is

PRACTICAL CONSIDERATIONS 93

repeated? Which variables can actually be measured before the outcome occurs? For example, if we want to predict the closing price of an ongoing online auc- tion, we cannot use the number of bids as a predictor because this will not be known until the auction closes.

Example 1: House Prices in Boston

We return to the Boston Housing example introduced in Chapter 3. For each neighborhood, a number of variables are given, such as the crime rate, the stu- dent/teacher ratio, and the median value of a housing unit in the neighborhood. A description of all 14 variables is given in Table 4.1. The first nine records of the data are shown in Table 4.2. The first row represents the first neighborhood, which had an average per capita crime rate of 0.006, 18% of the residential land zoned for lots over 25,000 ft2, 2.31% of the land devoted to nonretail business, no border on the Charles River, and so on.

TABLE 4.1 DESCRIPTION OF VARIABLES IN THE BOSTON HOUSING DATASET

CRIM Crime rate ZN Percentage of residential land zoned for lots over 25,000 ft2

INDUS Percentage of land occupied by nonretail business CHAS Does tract bound Charles River? (= 1 if tract bounds river, = 0 otherwise) NOX Nitric oxide concentration (parts per 10 million) RM Average number of rooms per dwelling AGE Percentage of owner-occupied units built prior to 1940 DIS Weighted distances to five Boston employment centers RAD Index of accessibility to radial highways TAX Full-value property tax rate per $10,000 PTRATIO Pupil-to-teacher ratio by town LSTAT Percentage of lower status of the population MEDV Median value of owner-occupied homes in $1000s CAT.MEDV Is median value of owner-occupied homes in tract above $30,000 (CAT.MEDV = 1) or

not (CAT.MEDV = 0)?

TABLE 4.2 FIRST NINE RECORDS IN THE BOSTON HOUSING DATA

CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT MEDV CAT.MEDV 0.00632 18.0 2.31 0 0.538 6.575 65.2 4.09 1 296 15.3 4.98 24.0 0

0.02731 0.0 7.07 0 0.469 6.421 78.9 4.9671 2 242 17.8 9.14 21.6 0

0.02729 0.0 7.07 0 0.469 7.185 61.1 4.9671 2 242 17.8 4.03 34.7 1

0.03237 0.0 2.18 0 0.458 6.998 45.8 6.0622 3 222 18.7 2.94 33.4 1

0.06905 0.0 2.18 0 0.458 7.147 54.2 6.0622 3 222 18.7 5.33 36.2 1

0.02985 0.0 2.18 0 0.458 6.43 58.7 6.0622 3 222 18.7 5.21 28.7 0

0.08829 12.5 7.87 0 0.524 6.012 66.6 5.5605 5 311 15.2 12.43 22.9 0

0.14455 12.5 7.87 0 0.524 6.172 96.1 5.9505 5 311 15.2 19.15 27.1 0

0.21124 12.5 7.87 0 0.524 5.631 100.0 6.0821 5 311 15.2 29.93 16.5 0

94 DIMENSION REDUCTION

4.4 Data Summaries

As we have seen in the chapter on data visualization, an important initial step of data exploration is getting familiar with the data and their characteristics through summaries and graphs. The importance of this step cannot be overstated. The better you understand the data, the better the results from the modeling or min- ing process.

Numerical summaries and graphs of the data are very helpful for data reduc- tion. The information that they convey can assist in combining categories of a categorical variable, in choosing variables to remove, in assessing the level of information overlap between variables, and more. Before discussing such strate- gies for reducing the dimension of a data set, let us consider useful summaries and tools.

Summary Statistics

R has several functions and facilities that assist in summarizing data. The func- tion summary() gives an overview of the entire set of variables in the data. The functions mean(), sd(), min(), max(), median(), and length() are also very helpful for learning about the characteristics of each variable. First, they give us infor- mation about the scale and type of values that the variable takes. The min and max functions can be used to detect extreme values that might be errors. The mean and median give a sense of the central values of that variable, and a large deviation between the two also indicates skew. The standard deviation gives a sense of how dispersed the data are (relative to the mean). Further options, such as sum(is.na(variable)), which gives the number of null values, can tell us about missing values.

Table 4.3 shows summary statistics (and R code) for the Boston Housing example. We immediately see that the different variables have very different ranges of values. We will soon see how variation in scale across variables can distort analyses if not treated properly. Another observation that can be made is that the mean of the first variable, CRIM (as well as several others), is much larger than the median, indicating right skew. None of the variables have missing values. There also do not appear to be indications of extreme values that might result from typing errors.

Next, we summarize relationships between two or more variables. For numerical variables, we can compute a complete matrix of correlations between each pair of variables, using the R function cor(). Table 4.4 shows the correlation matrix for the Boston Housing variables. We see that most correlations are low and that many are negative. Recall also the visual display of a correlation matrix via a heatmap (see Figure 3.4 in Chapter 3 for the heatmap corresponding to this

DATA SUMMARIES 95

TABLE 4.3 SUMMARY STATISTICS FOR THE BOSTON HOUSING DATA

code for summary statistics

boston.housing.df <- read.csv("BostonHousing.csv", header = True) head(boston.housing.df, 9) summary(boston.housing.df)

# compute mean, standard dev., min, max, median, length, and missing values of CRIM mean(boston.housing.df$CRIM) sd(boston.housing.df$CRIM) min(boston.housing.df$CRIM) max(boston.housing.df$CRIM) median(boston.housing.df$CRIM) length(boston.housing.df$CRIM)

# find the number of missing values of variable CRIM sum(is.na(boston.housing.df$CRIM))

# compute mean, standard dev., min, max, median, length, and missing values for all # variables data.frame(mean=sapply(boston.housing.df, mean), + + sd=sapply(boston.housing.df, sd), + + min=sapply(boston.housing.df, min), + + max=sapply(boston.housing.df, max), + + median=sapply(boston.housing.df, median), + + length=sapply(boston.housing.df, length) + + miss.val=sapply(boston.housing.df, function(x) + sum(length(which(is.na(x))))))) )

Output

> data.frame(mean=sapply(boston.housing.df, mean), + sd=sapply(boston.housing.df, sd), + min=sapply(boston.housing.df, min), + max=sapply(boston.housing.df, max), + median=sapply(boston.housing.df, median), + length=sapply(boston.housing.df, length) + miss.val=sapply(boston.housing.df,

function(x) sum(length(which(is.na(x))))))) mean sd min max median length miss.val

CRIM 3.61352356 8.6015451 0.00632 88.9762 0.25651 506 0 ZN 11.36363636 23.3224530 0.00000 100.0000 0.00000 506 0 INDUS 11.13677866 6.8603529 0.46000 27.7400 9.69000 506 0 CHAS 0.06916996 0.2539940 0.00000 1.0000 0.00000 506 0 NOX 0.55469506 0.1158777 0.38500 0.8710 0.53800 506 0 RM 6.28463439 0.7026171 3.56100 8.7800 6.20850 506 0 AGE 68.57490119 28.1488614 2.90000 100.0000 77.50000 506 0 DIS 3.79504269 2.1057101 1.12960 12.1265 3.20745 506 0 RAD 9.54940711 8.7072594 1.00000 24.0000 5.00000 506 0 TAX 408.23715415 168.5371161 187.00000 711.0000 330.00000 506 0 PTRATIO 18.45553360 2.1649455 12.60000 22.0000 19.05000 506 0 LSTAT 12.65306324 7.1410615 1.73000 37.9700 11.36000 506 0 MEDV 22.53280632 9.1971041 5.00000 50.0000 21.20000 506 0 CAT.MEDV 0.16600791 0.3724560 0.00000 1.0000 0.00000 506 0

96 DIMENSION REDUCTION

TABLE 4.4 CORRELATION TABLE FOR BOSTON HOUSING DATA

> round(cor(boston.housing.df),2) CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT MEDV CAT.MEDV

CRIM 1.00 -0.20 0.41 -0.06 0.42 -0.22 0.35 -0.38 0.63 0.58 0.29 0.46 -0.39 -0.15 ZN -0.20 1.00 -0.53 -0.04 -0.52 0.31 -0.57 0.66 -0.31 -0.31 -0.39 -0.41 0.36 0.37 INDUS 0.41 -0.53 1.00 0.06 0.76 -0.39 0.64 -0.71 0.60 0.72 0.38 0.60 -0.48 -0.37 CHAS -0.06 -0.04 0.06 1.00 0.09 0.09 0.09 -0.10 -0.01 -0.04 -0.12 -0.05 0.18 0.11 NOX 0.42 -0.52 0.76 0.09 1.00 -0.30 0.73 -0.77 0.61 0.67 0.19 0.59 -0.43 -0.23 RM -0.22 0.31 -0.39 0.09 -0.30 1.00 -0.24 0.21 -0.21 -0.29 -0.36 -0.61 0.70 0.64 AGE 0.35 -0.57 0.64 0.09 0.73 -0.24 1.00 -0.75 0.46 0.51 0.26 0.60 -0.38 -0.19 DIS -0.38 0.66 -0.71 -0.10 -0.77 0.21 -0.75 1.00 -0.49 -0.53 -0.23 -0.50 0.25 0.12 RAD 0.63 -0.31 0.60 -0.01 0.61 -0.21 0.46 -0.49 1.00 0.91 0.46 0.49 -0.38 -0.20 TAX 0.58 -0.31 0.72 -0.04 0.67 -0.29 0.51 -0.53 0.91 1.00 0.46 0.54 -0.47 -0.27 PTRATIO 0.29 -0.39 0.38 -0.12 0.19 -0.36 0.26 -0.23 0.46 0.46 1.00 0.37 -0.51 -0.44 LSTAT 0.46 -0.41 0.60 -0.05 0.59 -0.61 0.60 -0.50 0.49 0.54 0.37 1.00 -0.74 -0.47 MEDV -0.39 0.36 -0.48 0.18 -0.43 0.70 -0.38 0.25 -0.38 -0.47 -0.51 -0.74 1.00 0.79 CAT.MEDV -0.15 0.37 -0.37 0.11 -0.23 0.64 -0.19 0.12 -0.20 -0.27 -0.44 -0.47 0.79 1.00

correlation table). We will return to the importance of the correlation matrix soon, in the context of correlation analysis.

Aggregation and Pivot Tables

Another very useful approach for exploring the data is aggregation by one or more variables. For aggregation by a single variable, we can use table(). For example, Table 4.5 shows the number of neighborhoods that bound the Charles River vs. those that do not (the variable CHAS is chosen as the grouping vari- able). It appears that the majority of neighborhoods (471 of 506) do not bound the river.

The aggregate() function can be used for aggregating by one or more variables, and computing a range of summary statistics (count, average, percentage, etc.). For categorical variables, we obtain a breakdown of the records by the combina- tion of categories. For instance, in Table 4.6, we compute the average MEDV by CHAS and RM. Note that the numerical variable RM (the average number of rooms per dwelling in the neighborhood) should be first grouped into bins of size 1 (0–1, 1–2, and so on). Note the empty values, denoting that there are no neighborhoods in the dataset with those combinations (e.g., bounding the river and having on average 3 rooms).

TABLE 4.5 NUMBER OF NEIGHBORHOODS THAT BOUND THE CHARLES RIVER VS. THOSE THAT DO NOT

> boston.housing.df <- read.csv("BostonHousing.csv") > table(boston.housing.df$CHAS)

0 1 471 35

CORRELATION ANALYSIS 97

TABLE 4.6 AVERAGE MEDV BY CHAS AND RM

code for aggregating MEDV by CHAS and RM

# create bins of size 1 boston.housing.df$RM.bin <- .bincode(boston.housing.df$RM, c(1:9))

# compute the average of MEDV by (binned) RM and CHAS # in aggregate() use the argument by= to define the list of aggregating variables, # and FUN= as an aggregating function. aggregate(boston.housing.df$MEDV, by=list(RM=boston.housing.df$RM.bin,

CHAS=boston.housing.df$CHAS), FUN=mean)

Output

RM CHAS x 1 3 0 25.30000 2 4 0 15.40714 3 5 0 17.2000 4 6 0 21.76917 5 7 0 35.96444 6 8 0 45.70000 7 5 1 22.21818 8 6 1 25.91875 9 7 1 44.06667 10 8 1 35.95000

Another useful set of functions are melt() and cast() in the reshape package, that allow the creation of pivot tables. melt() takes a set of columns and stacks them into a single column. cast() then reshapes the single column into multiple columns by the aggregating variables of our choice. For example, Table 4.7 com- putes the average of MEDV by CHAS and RM and presents it as a pivot table.

In classification tasks, where the goal is to find predictor variables that dis- tinguish between two classes, a good exploratory step is to produce summaries for each class. This can assist in detecting useful predictors that display some separation between the two classes. Data summaries are useful for almost any data mining task and are therefore an important preliminary step for cleaning and understanding the data before carrying out further analyses.

4.5 Correlation Analysis

In datasets with a large number of variables (which are likely to serve as pre- dictors), there is usually much overlap in the information covered by the set of variables. One simple way to find redundancies is to look at a correlation matrix. This shows all the pairwise correlations between variables. Pairs that have a very strong (positive or negative) correlation contain a lot of overlap in information and are good candidates for data reduction by removing one of the variables.

98 DIMENSION REDUCTION

TABLE 4.7 PIVOT TABLES IN R

code for creating pivot tables using functions melt() and cast()

# use install.packages("reshape") the first time the package is used library(reshape) boston.housing.df <- read.csv("BostonHousing.csv") # create bins of size 1 boston.housing.df$RM.bin <- .bincode(boston.housing.df$RM, c(1:9))

# use melt() to stack a set of columns into a single column of data. # stack MEDV values for each combination of (binned) RM and CHAS mlt <- melt(boston.housing.df, id=c("RM.bin", "CHAS"), measure=c("MEDV")) head(mlt, 5)

# use cast() to reshape data and generate pivot table cast(mlt, RM.bin ~ CHAS, subset=variable=="MEDV",

margins=c("grand_row", "grand_col"), mean)

Output

> mlt <- melt(boston.housing.df, id=c("RM.bin", "CHAS"), measure=c("MEDV")) > head(mlt, 5)

RM.bin CHAS variable value 1 6 0 MEDV 24.0 2 6 0 MEDV 21.6 3 7 0 MEDV 34.7 4 6 0 MEDV 33.4 5 7 0 MEDV 36.2

> cast(mlt, RM.bin ~ CHAS, subset=variable=="MEDV", margins=c("grand_row", "grand_col"), mean)

RM.bin 0 1 (all) 1 3 25.30000 NaN 25.30000 2 4 15.40714 NaN 15.40714 3 5 17.20000 22.21818 17.55159 4 6 21.76917 25.91875 22.01599 5 7 35.96444 44.06667 36.91765 6 8 45.70000 35.95000 44.20000 7 (all) 22.09384 28.44000 22.53281

Removing variables that are strongly correlated to others is useful for avoiding multicollinearity problems that can arise in various models. (Multicollinearity is the presence of two or more predictors sharing the same linear relationship with the outcome variable; R handles this automatically in regression.)

Correlation analysis is also a good method for detecting duplications of vari- ables in the data. Sometimes, the same variable appears accidentally more than once in the dataset (under a different name) because the dataset was merged from multiple sources, the same phenomenon is measured in different units, and so

REDUCING THE NUMBER OF CATEGORIES IN CATEGORICAL VARIABLES 99

on. Using correlation table heatmaps, as shown in Chapter 3, can make the task of identifying strong correlations easier.

4.6 Reducing the Number of Categories in Categorical Variables

When a categorical variable has many categories, and this variable is destined to be a predictor, many data mining methods will require converting it into many dummy variables. In particular, a variable with m categories will be trans- formed into either m or m − 1 dummy variables (depending on the method). This means that even if we have very few original categorical variables, they can greatly inflate the dimension of the dataset. One way to handle this is to reduce the number of categories by combining close or similar categories. Combining categories requires incorporating expert knowledge and common sense. Pivot tables are useful for this task: We can examine the sizes of the various categories and how the outcome variable behaves in each category. Generally, categories that contain very few observations are good candidates for combining with other categories. Use only the categories that are most relevant to the analysis and label the rest as “other.” In classification tasks (with a categorical outcome variable), a pivot table broken down by the outcome classes can help identify categories that do not separate the classes. Those categories too are candidates for inclusion in the “other” category. An example is shown in Figure 4.1, where the distribution of outcome variable CAT.MEDV is broken down by ZN (treated here as a cate- gorical variable). We can see that the distribution of CAT.MEDV is identical for ZN = 17.5, 90, 95, and 100 (where all neighborhoods have CAT.MEDV = 1). These four categories can then be combined into a single category. Similarly, categories ZN = 12.5, 25, 28, 30, and 70 can be combined. Further combina- tion is also possible based on similar bars.

In a time series context where we might have a categorical variable denoting season (such as month, or hour of day) that will serve as a predictor, reducing categories can be done by examining the time series plot and identifying similar periods. For example, the time plot in Figure 4.2 shows the quarterly revenues of Toys “R” Us between 1992–1995. Only quarter 4 periods appear different, and therefore, we can combine quarters 1–3 into a single category.

4.7 Converting a Categorical Variable to a Numerical Variable

Sometimes the categories in a categorical variable represent intervals. Common examples are age group or income bracket. If the interval values are known (e.g., category 2 is the age interval 20–30), we can replace the categorical value

100 DIMENSION REDUCTION

0 17.5 20 22 28 33 35 45 55 70 80 85 95

Distribution of CAT.MEDV by ZN

ZN

0%

20 %

40 %

60 %

80 %

10 0%

FIGURE 4.1 DISTRIBUTION OF CAT.MEDV (BLACK DENOTES CAT.MEDV = 0) BY ZN. SIMILAR BARS INDICATE LOW SEPARATION BETWEEN CLASSES, AND CAN BE COMBINED

code for creating Figure 4.1

library(ggmap) boston.housing.df <- read.csv("BostonHousing.csv")

tbl <- table(boston.housing.df$CAT..MEDV, boston.housing.df$ZN) prop.tbl <- prop.table(tbl, margin=2) barplot(prop.tbl, xlab="ZN", ylab="", yaxt="n",main="Distribution of CAT.MEDV by ZN") axis(2, at=(seq(0,1, 0.2)), paste(seq(0,100,20), "%"))

Time

R ev

en ue

( $

m ill

io ns

)

1992 1993 1994 1995 1996

10 00

15 00

20 00

25 00

30 00

35 00

40 00

FIGURE 4.2 QUARTERLY REVENUES OF TOYS “R” US, 1992–1995

PRINCIPAL COMPONENTS ANALYSIS 101

(“2” in the example) with the mid-interval value (here “25”). The result will be a numerical variable which no longer requires multiple dummy variables.

4.8 Principal Components Analysis

Principal components analysis (PCA) is a useful method for dimension reduction, especially when the number of variables is large. PCA is especially valuable when we have subsets of measurements that are measured on the same scale and are highly correlated. In that case, it provides a few variables (often as few as three) that are weighted linear combinations of the original variables, and that retain the majority of the information of the full original set. PCA is intended for use with numerical variables. For categorical variables, other methods such as correspondence analysis are more suitable.

Example 2: Breakfast Cereals

Data were collected on the nutritional information and consumer rating of 77 breakfast cereals.1 The consumer rating is a rating of cereal “healthiness” for consumer information (not a rating by consumers). For each cereal, the data include 13 numerical variables, and we are interested in reducing this dimen- sion. For each cereal, the information is based on a bowl of cereal rather than a serving size, because most people simply fill a cereal bowl (resulting in constant volume, but not weight). A snapshot of these data is given in Figure 4.3, and the description of the different variables is given in Table 4.8.

We focus first on two variables: calories and consumer rating. These are given in Table 4.9. The average calories across the 77 cereals is 106.88 and the average consumer rating is 42.67. The estimated covariance matrix between the two variables is

S =

[ 379.63 −188.68

−188.68 197.32

] .

It can be seen that the two variables are strongly correlated with a negative correlation of

−0.69 = −188.68√ (379.63)(197.32)

.

Roughly speaking, 69% of the total variation in both variables is actually “co- variation,” or variation in one variable that is duplicated by similar variation in the other variable. Can we use this fact to reduce the number of variables, while making maximum use of their unique contributions to the overall variation? Since there is redundancy in the information that the two variables contain, it

1The data are available at http://lib.stat.cmu.edu/DASL/Stories/HealthyBreakfast.html.

102 DIMENSION REDUCTION

Cereal Name mfr type calories protein fat sodium fiber carbo sugars potass vitamins 100% Bran N C 70 4 1 130 10 5 6 280 25 100% Natural Bran Q C 120 3 5 15 2 8 8 135 0 All-Bran K C 70 4 1 260 9 7 5 320 25 All-Bran with Extra Fiber K C 50 4 0 140 14 8 0 330 25 Almond Delight R C 110 2 2 200 1 14 8 25 Apple Cinnamon Cheerios G C 110 2 2 180 1.5 10.5 10 70 25 Apple Jacks K C 110 2 0 125 1 11 14 30 25 Basic 4 G C 130 3 2 210 2 18 8 100 25 Bran Chex R C 90 2 1 200 4 15 6 125 25 Bran Flakes P C 90 3 0 210 5 13 5 190 25 Cap'n'Crunch Q C 120 1 2 220 0 12 12 35 25 Cheerios G C 110 6 2 290 2 17 1 105 25 Cinnamon Toast Crunch G C 120 1 3 210 0 13 9 45 25 Clusters G C 110 3 2 140 2 13 7 105 25 Cocoa Puffs G C 110 1 1 180 0 12 13 55 25 Corn Chex R C 110 2 0 280 0 22 3 25 25 Corn Flakes K C 100 2 0 290 1 21 2 35 25 Corn Pops K C 110 1 0 90 1 13 12 20 25 Count Chocula G C 110 1 1 180 0 12 13 65 25 Cracklin' Oat Bran K C 110 3 3 140 4 10 7 160 25

FIGURE 4.3 SAMPLE FROM THE 77 BREAKFAST CEREALS DATASET

might be possible to reduce the two variables to a single variable without losing too much information. The idea in PCA is to find a linear combination of the two variables that contains most, even if not all, of the information, so that this new variable can replace the two original variables. Information here is in the sense of variability: What can explain the most variability among the 77 cereals? The total variability here is the sum of the variances of the two variables, which

TABLE 4.8 DESCRIPTION OF THE VARIABLES IN THE BREAKFAST CEREAL DATASET

Variable Description

mfr Manufacturer of cereal (American Home Food Products, General Mills, Kellogg, etc.)

type Cold or hot

calories Calories per serving

protein Grams of protein

fat Grams of fat

sodium Milligrams of sodium

fiber Grams of dietary fiber

carbo Grams of complex carbohydrates

sugars Grams of sugars

potass Milligrams of potassium

vitamins Vitamins and minerals: 0, 25, or 100, indicating the typical percentage of FDA recommended

shelf Display shelf (1, 2, or 3, counting from the floor)

weight Weight in ounces of one serving

cups Number of cups in one serving

rating Rating of the cereal calculated by consumer reports

PRINCIPAL COMPONENTS ANALYSIS 103

TABLE 4.9 CEREAL CALORIES AND RATINGS

Cereal Calories Rating Cereal Calories Rating

100% Bran 70 68.40297 Just Right Fruit & Nut 140 36.471512

100% Natural Bran 120 33.98368 Kix 110 39.241114

All-Bran 70 59.42551 Life 100 45.328074

All-Bran with Extra Fiber

50 93.70491 Lucky Charms 110 26.734515

Almond Delight 110 34.38484 Maypo 100 54.850917

Apple Cinnamon Cheerios

110 29.50954 Muesli Raisins, Dates & Almonds

150 37.136863

Apple Jacks 110 33.17409 Muesli Raisins, Peaches & Pecans

150 34.139765

Basic 4 130 37.03856 Mueslix Crispy Blend 160 30.313351

Bran Chex 90 49.12025 Multi-Grain Cheerios 100 40.105965

Bran Flakes 90 53.31381 Nut&Honey Crunch 120 29.924285

Cap’n’Crunch 120 18.04285 Nutri-Grain Almond-Raisin

140 40.69232

Cheerios 110 50.765 Nutri-grain Wheat 90 59.642837

Cinnamon Toast Crunch

120 19.82357 Oatmeal Raisin Crisp 130 30.450843

Clusters 110 40.40021 Post Nat. Raisin Bran 120 37.840594

Cocoa Puffs 110 22.73645 Product 19 100 41.50354

Corn Chex 110 41.44502 Puffed Rice 50 60.756112

Corn Flakes 100 45.86332 Puffed Wheat 50 63.005645

Corn Pops 110 35.78279 Quaker Oat Squares 100 49.511874

Count Chocula 110 22.39651 Quaker Oatmeal 100 50.828392

Cracklin’ Oat Bran 110 40.44877 Raisin Bran 120 39.259197

Cream of Wheat (Quick) 100 64.53382 Raisin Nut Bran 100 39.7034

Crispix 110 46.89564 Raisin Squares 90 55.333142

Crispy Wheat & Raisins 100 36.1762 Rice Chex 110 41.998933

Double Chex 100 44.33086 Rice Krispies 110 40.560159

Froot Loops 110 32.20758 Shredded Wheat 80 68.235885

Frosted Flakes 110 31.43597 Shredded Wheat ’n’Bran

90 74.472949

Frosted Mini-Wheats 100 58.34514 Shredded Wheat spoon size

90 72.801787

Fruit & Fibre Dates, Walnuts & Oats

120 40.91705 Smacks 110 31.230054

Fruitful Bran 120 41.01549 Special K 110 53.131324

Fruity Pebbles 110 28.02577 Strawberry Fruit Wheats

90 59.363993

Golden Crisp 100 35.25244 Total Corn Flakes 110 38.839746

Golden Grahams 110 23.80404 Total Raisin Bran 140 28.592785

Grape Nuts Flakes 100 52.0769 Total Whole Grain 100 46.658844

Grape-Nuts 110 53.37101 Triples 110 39.106174

Great Grains Pecan 120 45.81172 Trix 110 27.753301

Honey Graham Ohs 120 21.87129 Wheat Chex 100 49.787445

Honey Nut Cheerios 110 31.07222 Wheaties 100 51.592193

Honey-comb 110 28.74241 Wheaties Honey Gold 110 36.187559

Just Right Crunchy Nuggets

110 36.52368

104 DIMENSION REDUCTION

in this case is 379.63 + 197.32 = 577. This means that calories accounts for 66% = 379.63 / 577 of the total variability, and rating for the remaining 34%. If we drop one of the variables for the sake of dimension reduction, we lose at least 34% of the total variability. Can we redistribute the total variability between two new variables in a more polarized way? If so, it might be possible to keep only the one new variable that (hopefully) accounts for a large portion of the total variation.

Figure 4.4 shows a scatter plot of rating vs. calories. The line z1 is the direction in which the variability of the points is largest. It is the line that captures the most variation in the data if we decide to reduce the dimensionality of the data from two to one. Among all possible lines, it is the line for which, if we project the points in the dataset orthogonally to get a set of 77 (one-dimensional) values, the variance of the z1 values will be maximum. This is called the first principal component. It is also the line that minimizes the sum-of-squared perpendicular distances from the line. The z2-axis is chosen to be perpendicular to the z1-axis. In the case of two variables, there is only one line that is perpendicular to z1, and it has the second largest variability, but its information is uncorrelated with z1. This is called the second principal component. In general, when we have more than two variables, once we find the direction z1 with the largest variability,

FIGURE 4.4 SCATTER PLOT OF RATING VS. CALORIES FOR 77 BREAKFAST CEREALS, WITH THE TWO PRINCIPAL COMPONENT DIRECTIONS

PRINCIPAL COMPONENTS ANALYSIS 105

we search among all the orthogonal directions to z1 for the one with the next- highest variability. That is z2. The idea is then to find the coordinates of these lines and to see how they redistribute the variability.

Running PCA in R is done with the function prcomp(). Table 4.10 shows the output from running PCA on the two variables calories and rating. The value .rot of this function is the rotation matrix, which gives the weights that are used to project the original points onto the two new directions. The weights for z1 are given by (0.847, −0.532), and for z2 they are given by (0.532, 0.847). The function summary() gives the reallocated variance: z1 accounts for 86% of the total variability and z2 for the remaining 14%. Therefore, if we drop z2, we still maintain 86% of the total variability.

TABLE 4.10 PCA ON THE TWO VARIABLES CALORIES AND RATING

code for running PCA

cereals.df <- read.csv("Cereals.csv") # compute PCs on two dimensions pcs <- prcomp(data.frame(cereals.df$calories, cereals.df$rating)) summary(pcs) pcs$rot scores <- pcs$x head(scores, 5)

Output

> summary(pcs)

Importance of components: PC1 PC2

Standard deviation 22.3165 8.8844 Proportion of Variance 0.8632 0.1368 Cumulative Proportion 0.8632 1.0000

> pcs$rot

PC1 PC2 cereals.df.calories 0.8470535 0.5315077 cereals.df.rating -0.5315077 0.8470535

> scores <- pcs$x > head(scores, 5)

PC1 PC2 [1,] -44.921528 2.1971833 [2,] 15.725265 -0.3824165 [3,] -40.149935 -5.4072123 [4,] -75.310772 12.9991256 [5,] 7.041508 -5.3576857

106 DIMENSION REDUCTION

The weights are used to compute principal component scores, which are the projected values of calories and rating onto the new axes (after subtracting the means). In R, scores for the two dimensions are depicted by the PCA value .x. The first column is the projection onto z1 using the weights (0.847, −0.532). The second column is the projection onto z2 using the weights (0.532, 0.847). For example, the first score for the 100% Bran cereal (with 70 calories and a rating of 68.4) is (0.847)(70 − 106.88) + (−0.532)(68.4 − 42.67) = −44.92.

Note that the means of the new variables z1 and z2 are zero, because we’ve subtracted the mean of each variable. The sum of the variances var(z1) + var(z2) is equal to the sum of the variances of the original variables, var(calories)+var(rating). Furthermore, the variances of z1 and z2 are 498 and 79, respectively, so the first principal component, z1, accounts for 86% of the total variance. Since it captures most of the variability in the data, it seems reasonable to use one variable, the first principal score, to represent the two variables in the original data. Next, we generalize these ideas to more than two variables.

Principal Components

Let us formalize the procedure described above so that it can easily be generalized to p > 2 variables. Denote the original p variables by X1, X2, . . . , Xp. In PCA, we are looking for a set of new variables Z1, Z2, . . . , Zp that are weighted averages of the original variables (after subtracting their mean):

Zi = ai,1(X1 − X̄1) + ai,2(X2 − X̄2) + · · ·+ ai,p(Xp − X̄p), i = 1, . . . , p (4.1)

where each pair of Z’s has correlation = 0. We then order the resulting Z’s by their variance, with Z1 having the largest variance and Zp having the small- est variance. The software computes the weights ai,j , which are then used in computing the principal component scores.

A further advantage of the principal components compared to the original data is that they are uncorrelated (correlation coefficient = 0). If we construct regression models using these principal components as predictors, we will not encounter problems of multicollinearity.

Let us return to the breakfast cereal dataset with all 15 variables, and apply PCA to the 13 numerical variables. The resulting output is shown in Table 4.11. Note that the first three components account for more than 96% of the total variation associated with all 13 of the original variables. This suggests that we can capture most of the variability in the data with less than 25% of the original dimensions in the data. In fact, the first two principal components alone capture 92.6% of the total variation. However, these results are influenced by the scales of the variables, as we describe next.

PRINCIPAL COMPONENTS ANALYSIS 107

TABLE 4.11 PCA OUTPUT USING ALL 13 NUMERICAL VARIABLES IN THE BREAKFAST CEREALS DATASET. THE TABLE SHOWS RESULTS FOR THE FIRST FIVE PRINCIPAL COMPONENTS

> pcs <- prcomp(na.omit(cereals.df[,-c(1:3)])) > summary(pcs)

Importance of components: PC1 PC2 PC3 PC4 PC5 PC6 PC7

Standard deviation 83.7641 70.9143 22.64375 19.18148 8.42323 2.09167 1.69942 Proportion of Variance 0.5395 0.3867 0.03943 0.02829 0.00546 0.00034 0.00022 Cumulative Proportion 0.5395 0.9262 0.96560 0.99389 0.99935 0.99968 0.99991

PC8 PC9 PC10 PC11 PC12 PC13 Standard deviation 0.77963 0.65783 0.37043 0.1864 0.06302 5.334e-08 Proportion of Variance 0.00005 0.00003 0.00001 0.0000 0.00000 0.000e+00 Cumulative Proportion 0.99995 0.99999 1.00000 1.0000 1.00000 1.000e+00

> pcs$rot[,1:5]

PC1 PC2 PC3 PC4 PC5 calories 0.0779841812 0.0093115874 -0.6292057595 -0.6010214629 0.454958508 protein -0.0007567806 -0.0088010282 -0.0010261160 0.0031999095 0.056175970 fat -0.0001017834 -0.0026991522 -0.0161957859 -0.0252622140 -0.016098458 sodium 0.9802145422 -0.1408957901 0.1359018583 -0.0009680741 0.013948118 fiber -0.0054127550 -0.0306807512 0.0181910456 0.0204721894 0.013605026 carbo 0.0172462607 0.0167832981 -0.0173699816 0.0259482087 0.349266966 sugars 0.0029888631 0.0002534853 -0.0977049979 -0.1154809105 -0.299066459 potass -0.1349000039 -0.9865619808 -0.0367824989 -0.0421757390 -0.047150529 vitamins 0.0942933187 -0.0167288404 -0.6919777623 0.7141179984 -0.037008623 shelf -0.0015414195 -0.0043603994 -0.0124888415 0.0056471836 -0.007876459 weight 0.0005120017 -0.0009992138 -0.0038059565 -0.0025464145 0.003022113 cups 0.0005101111 0.0015910125 -0.0006943214 0.0009853800 0.002148458 rating -0.0752962922 -0.0717421528 0.3079471212 0.3345338994 0.757708025

comment on missing values

Use function na.omit() to remove observations that contain missing values.

Normalizing the Data

A further use of PCA is to understand the structure of the data. This is done by examining the weights to see how the original variables contribute to the different principal components. In our example, it is clear that the first principal component is dominated by the sodium content of the cereal: it has the highest (in this case, positive) weight. This means that the first principal component is in fact measuring how much sodium is in the cereal. Similarly, the second principal component seems to be measuring the amount of potassium. Since both these variables are measured in milligrams, whereas the other nutrients are measured in grams, the scale is obviously leading to this result. The variances of potassium and sodium are much larger than the variances of the other variables, and thus the total variance is dominated by these two variances. A solution is to normalize the data before performing the PCA. Normalization (or standardization) means replacing each original variable by a standardized version of the variable that

108 DIMENSION REDUCTION

has unit variance. This is easily accomplished by dividing each variable by its standard deviation. The effect of this normalization is to give all variables equal importance in terms of variability.

When should we normalize the data like this? It depends on the nature of the data. When the units of measurement are common for the variables (e.g., dollars), and when their scale reflects their importance (sales of jet fuel, sales of heating oil), it is probably best not to normalize (i.e., not to rescale the data so that they have unit variance). If the variables are measured in different units so that it is unclear how to compare the variability of different variables (e.g., dollars for some, parts per million for others) or if for variables measured in the same units, scale does not reflect importance (earnings per share, gross revenues), it is generally advisable to normalize. In this way, the differences in units of measurement do not affect the principal components’ weights. In the rare situations where we can give relative weights to variables, we multiply the normalized variables by these weights before doing the principal components analysis.

Thus far, we have calculated principal components using the covariance matrix. An alternative to normalizing and then performing PCA is to perform PCA on the correlation matrix instead of the covariance matrix. Most software programs allow the user to choose between the two. Remember that using the correlation matrix means that you are operating on the normalized data.

Returning to the breakfast cereal data, we normalize the 13 variables due to the different scales of the variables and then perform PCA (or equiva- lently, we use PCA applied to the correlation matrix). The output is given in Table 4.12.

Now we find that we need 7 principal components to account for more than 90% of the total variability. The first 2 principal components account for only 52% of the total variability, and thus reducing the number of variables to two would mean losing a lot of information. Examining the weights, we see that the first principal component measures the balance between 2 quantities: (1) calories and cups (large positive weights) vs. (2) protein, fiber, potassium, and consumer rating (large negative weights). High scores on principal component 1 mean that the cereal is high in calories and the amount per bowl, and low in protein, and potassium. Unsurprisingly, this type of cereal is associated with a low consumer rating. The second principal component is most affected by the weight of a serving, and the third principal component by the carbohydrate content. We can continue labeling the next principal components in a similar fashion to learn about the structure of the data.

When the data can be reduced to two dimensions, a useful plot is a scatter plot of the first vs. second principal scores with labels for the observations (if

PRINCIPAL COMPONENTS ANALYSIS 109

TABLE 4.12 PCA OUTPUT USING ALL NORMALIZED 13 NUMERICAL VARIABLES IN THE BREAKFAST CEREALS DATASET. THE TABLE SHOWS RESULTS FOR THE FIRST FIVE PRINCIPAL COMPONENTS

> pcs.cor <- prcomp(na.omit(cereals.df[,-c(1:3)]), scale. = T) > summary(pcs.cor)

Importance of components: PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8

Standard deviation 1.9062 1.7743 1.3818 1.00969 0.9947 0.84974 0.81946 0.64515 Proportion of Variance 0.2795 0.2422 0.1469 0.07842 0.0761 0.05554 0.05166 0.03202 Cumulative Proportion 0.2795 0.5217 0.6685 0.74696 0.8231 0.87861 0.93026 0.96228

PC9 PC10 PC11 PC12 PC13 Standard deviation 0.56192 0.30301 0.25194 0.13897 1.499e-08 Proportion of Variance 0.02429 0.00706 0.00488 0.00149 0.000e+00 Cumulative Proportion 0.98657 0.99363 0.99851 1.00000 1.000e+00

> pcs.cor$rot[,1:5]

PC1 PC2 PC3 PC4 PC5 calories 0.29954236 0.3931479 -0.114857453 0.20435870 0.20389885 protein -0.30735632 0.1653233 -0.277281953 0.30074318 0.31974897 fat 0.03991542 0.3457243 0.204890102 0.18683311 0.58689327 sodium 0.18339651 0.1372205 -0.389431009 0.12033726 -0.33836424 fiber -0.45349036 0.1798119 -0.069766079 0.03917361 -0.25511906 carbo 0.19244902 -0.1494483 -0.562452458 0.08783547 0.18274252 sugars 0.22806849 0.3514345 0.355405174 -0.02270716 -0.31487243 potass -0.40196429 0.3005442 -0.067620183 0.09087843 -0.14836048 vitamins 0.11598020 0.1729092 -0.387858660 -0.60411064 -0.04928672 shelf -0.17126336 0.2650503 0.001531036 -0.63887859 0.32910135 weight 0.05029930 0.4503085 -0.247138314 0.15342874 -0.22128334 cups 0.29463553 -0.2122479 -0.139999705 0.04748909 0.12081645 rating -0.43837841 -0.2515389 -0.181842433 0.03831622 0.05758420

comment on normalization

Use function prcomp() with scale. = T to run PCA on normalized data.

the dataset is not too large). To illustrate this, Figure 4.5 displays the first two principal component scores for the breakfast cereals.

We can see that as we move from right (bran cereals) to left, the cereals are less “healthy” in the sense of high calories, low protein and fiber, and so on. Also, moving from bottom to top, we get heavier cereals (moving from puffed rice to raisin bran). These plots are especially useful if interesting clusters of observations can be found. For instance, we see here that children’s cereals are close together on the middle-left part of the plot.

Using Principal Components for Classification and Prediction

When the goal of the data reduction is to have a smaller set of variables that will serve as predictors, we can proceed as following: Apply PCA to the predictors

110 DIMENSION REDUCTION

FIGURE 4.5 SCATTER PLOT OF THE SECOND VS. FIRST PRINCIPAL COMPONENTS SCORES FOR THE NORMALIZED BREAKFAST CEREAL OUTPUT (CREATED USING TABLEAU)

using the training data. Use the output to determine the number of principal components to be retained. The predictors in the model now use the (reduced number of) principal scores columns. For the validation set, we can use the weights computed from the training data to obtain a set of principal scores by applying the weights to the variables in the validation set. These new variables are then treated as the predictors.

One disadvantage of using a subset of principal components as predictors in a supervised task, is that we might lose predictive information that is nonlinear

DIMENSION REDUCTION USING REGRESSION MODELS 111

(e.g., a quadratic effect of a predictor on the outcome variable or an interac- tion between predictors). This is because PCA produces linear transformations, thereby capturing linear relationships between the original variables.

4.9 Dimension Reduction Using Regression Models

In this chapter, we discussed methods for reducing the number of columns using summary statistics, plots, and principal components analysis. All these are con- sidered exploratory methods. Some of them completely ignore the outcome variable (e.g., PCA), whereas in other methods we informally try to incorporate the relationship between the predictors and the outcome variable (e.g., combin- ing similar categories, in terms of their outcome variable behavior). Another approach to reducing the number of predictors, which directly considers the predictive or classification task, is by fitting a regression model. For prediction, a linear regression model is used (see Chapter 6) and for classification, a logistic regression model (see Chapter 10). In both cases, we can employ subset selection procedures that algorithmically choose a subset of predictor variables among the larger set (see details in the relevant chapters).

Fitted regression models can also be used to further combine similar cate- gories: categories that have coefficients that are not statistically significant (i.e., have a high p-value) can be combined with the reference category, because their distinction from the reference category appears to have no significant effect on the outcome variable. Moreover, categories that have similar coefficient values (and the same sign) can often be combined, because their effect on the outcome variable is similar. See the example in Chapter 10 on predicting delayed flights for an illustration of how regression models can be used for dimension reduction.

4.10 Dimension Reduction Using Classification and Regression Trees

Another method for reducing the number of columns and for combining cat- egories of a categorical variable is by applying classification and regression trees (see Chapter 9). Classification trees are used for classification tasks and regression trees for prediction tasks. In both cases, the algorithm creates binary splits on the predictors that best classify/predict the outcome variable (e.g., above/below age 30). Although we defer the detailed discussion to Chapter 9, we note here that the resulting tree diagram can be used for determining the important predic- tors. Predictors (numerical or categorical) that do not appear in the tree can be removed. Similarly, categories that do not appear in the tree can be combined.

112 DIMENSION REDUCTION

PROBLEMS

4.1 Breakfast Cereals. Use the data for the breakfast cereals example in Section 4.8 to explore and summarize the data as follows:

a. Which variables are quantitative/numerical? Which are ordinal? Which are nominal?

b. Compute the mean, median, min, max, and standard deviation for each of the quantitative variables. This can be done through R’s sapply() function (e.g., sap- ply(data, mean, na.rm = TRUE)).

c. Use R to plot a histogram for each of the quantitative variables. Based on the histograms and summary statistics, answer the following questions:

i. Which variables have the largest variability?

ii. Which variables seem skewed?

iii. Are there any values that seem extreme?

d. Use R to plot a side-by-side boxplot comparing the calories in hot vs. cold cereals. What does this plot show us?

e. Use R to plot a side-by-side boxplot of consumer rating as a function of the shelf height. If we were to predict consumer rating from shelf height, does it appear that we need to keep all three categories of shelf height?

f. Compute the correlation table for the quantitative variable (function cor()). In addi- tion, generate a matrix plot for these variables (function plot(data)).

i. Which pair of variables is most strongly correlated?

ii. How can we reduce the number of variables based on these correlations?

iii. How would the correlations change if we normalized the data first?

g. Consider the first PC of the analysis of the 13 numerical variables in Table 4.11. Describe briefly what this PC represents.

4.2 University Rankings. The dataset on American college and university rankings (available from www.dataminingbook.com) contains information on 1302 American colleges and universities offering an undergraduate program. For each university, there are 17 measurements that include continuous measurements (such as tuition and grad- uation rate) and categorical measurements (such as location by state and whether it is a private or a public school).

a. Remove all categorical variables. Then remove all records with missing numerical measurements from the dataset.

b. Conduct a principal components analysis on the cleaned data and comment on the results. Should the data be normalized? Discuss what characterizes the components you consider key.

4.3 Sales of Toyota Corolla Cars. The file ToyotaCorolla.csv contains data on used cars (Toyota Corollas) on sale during late summer of 2004 in the Netherlands. It has 1436 records containing details on 38 attributes, including Price, Age, Kilometers, HP, and other specifications. The goal will be to predict the price of a used Toyota Corolla based on its specifications.

a. Identify the categorical variables.

b. Explain the relationship between a categorical variable and the series of binary dummy variables derived from it.

PROBLEMS 113

c. How many dummy binary variables are required to capture the information in a categorical variable with N categories?

d. Use R to convert the categorical variables in this dataset into dummy variables, and explain in words, for one record, the values in the derived binary dummies.

e. Use R to produce a correlation matrix and matrix plot. Comment on the relation- ships among variables.

4.4 Chemical Features of Wine. Table 4.13 shows the PCA output on data (non- normalized) in which the variables represent chemical characteristics of wine, and each case is a different wine.

TABLE 4.13 PRINCIPAL COMPONENTS OF NON-NORMALIZED WINE DATA

code for running PCA on the wine data

wine.df <- read.csv("Wine.csv") pcs.cor <- prcomp(wine.df[,-1]) summary(pcs.cor) pcs.cor$rot[,1:4]

Output

> summary(pcs.cor)

importance of components: PC1 PC2 PC3 PC4 PC5

Standard deviation 314.9632 13.13527 3.07215 2.23409 1.10853 Proportion of Variance 0.9981 0.00174 0.00009 0.00005 0.00001 Cumulative Proportion 0.9981 0.99983 0.99992 0.99997 0.99998

PC6 PC7 PC8 PC9 PC10 Standard deviation 0.91710 0.5282 0.3891 0.3348 0.2678 Proportion of Variance 0.00001 0.0000 0.0000 0.0000 0.0000 Cumulative Proportion 0.99999 1.0000 1.0000 1.0000 1.0000

PC11 PC12 PC13 Standard deviation 0.1938 0.1452 0.09057 Proportion of Variance 0.0000 0.0000 0.00000 Cumulative Proportion 1.0000 1.0000 1.00000

> pcs.cor$rot[,1:4]

PC1 PC2 PC3 PC4 Alcohol -0.0016592647 -1.203406e-03 -0.016873809 0.141446778 Malic_Acid 0.0006810156 -2.154982e-03 -0.122003373 0.160389543 Ash -0.0001949057 -4.593693e-03 -0.051987430 -0.009772810 Ash_Alcalinity 0.0046713006 -2.645039e-02 -0.938593003 -0.330965260 Magnesium -0.0178680075 -9.993442e-01 0.029780248 -0.005393756 Total_Phenols -0.0009898297 -8.779622e-04 0.040484644 -0.074584656 Flavanoids -0.0015672883 5.185073e-05 0.085443339 -0.169086724 Nonflavanoid_Phenols 0.0001230867 1.354479e-03 -0.013510780 0.010805561 Proanthocyanins -0.0006006078 -5.004400e-03 0.024659382 -0.050120952 Color_Intensity -0.0023271432 -1.510035e-02 -0.291398464 0.878893693 Hue -0.0001713800 7.626731e-04 0.025977662 -0.060034945 OD280_OD315 -0.0007049316 3.495364e-03 0.070323969 -0.178200254 Proline -0.9998229365 1.777381e-02 -0.004528682 -0.003112916

114 DIMENSION REDUCTION

a. The data are in the file Wine.csv. Consider the rows labeled “Proportion of Vari- ance.” Explain why the value for PC1 is so much greater than that of any other column.

b. Comment on the use of normalization (standardization) in part (a).

Part III

Performance Evaluation

CHAPTER 5

Evaluating Predictive Performance

In this chapter, we discuss how the predictive performance of data mining meth- ods can be assessed. We point out the danger of overfitting to the training data, and the need to test model performance on data that were not used in the training step. We discuss popular performance metrics. For prediction, metrics include Average Error, MAPE, and RMSE (based on the validation data). For classi- fication tasks, metrics based on the confusion matrix include overall accuracy, specificity and sensitivity, and metrics that account for misclassification costs. We also show the relation between the choice of cutoff value and classification performance, and present the ROC curve, which is a popular chart for assessing method performance at different cutoff values. When the goal is to accurately classify the most interesting or important records, called ranking, rather than accu- rately classify the entire sample (e.g., the 10% of customers most likely to respond to an offer, or the 5% of claims most likely to be fraudulent), lift charts are used to assess performance. We also discuss the need for oversampling rare classes and how to adjust performance metrics for the oversampling. Finally, we men- tion the usefulness of comparing metrics based on the validation data to those based on the training data for the purpose of detecting overfitting. While some differences are expected, extreme differences can be indicative of overfitting.

5.1 Introduction

In supervised learning, we are interested in predicting the outcome variable for new records. Three main types of outcomes of interest are:

Predicted numerical value : when the outcome variable is numerical (e.g., house price)

117

118 EVALUATING PREDICTIVE PERFORMANCE

Predicted class membership : when the outcome variable is categorical (e.g., buyer/nonbuyer)

Propensity : the probability of class membership, when the outcome vari- able is categorical (e.g., the propensity to default)

Prediction methods are used for generating numerical predictions, while classi- fication methods (“classifiers”) are used for generating propensities and, using a cutoff value on the propensities, we can generate predicted class memberships.

A subtle distinction to keep in mind is the two distinct predictive uses of classifiers: one use, classification, is aimed at predicting class membership for new records. The other, ranking, is detecting among a set of new records the ones most likely to belong to a class of interest.

Let’s now examine the approach for judging the usefulness of a prediction method used for generating numerical predictions (Section 5.2), a classifier used for classification (Section 5.3), and a classifier used for ranking (Section 5.4). In Section 5.5, we’ll look at evaluating performance under the scenario of over- sampling.

5.2 Evaluating Predictive Performance

First, let us emphasize that predictive accuracy is not the same as goodness-of- fit. Classical statistical measures of performance are aimed at finding a model that fits well to the data on which the model was trained. In data mining, we are interested in models that have high predictive accuracy when applied to new records. Measures such as R2 and standard error of estimate are common metrics in classical regression modeling, and residual analysis is used to gauge goodness- of-fit in that situation. However, these measures do not tell us much about the ability of the model to predict new records.

For assessing prediction performance, several measures are used. In all cases, the measures are based on the validation set, which serves as a more objective ground than the training set to assess predictive accuracy. This is because records in the validation set are more similar to the future records to be predicted, in the sense that they are not used to select predictors or to estimate the model parameters. Models are trained on the training data, applied to the validation data, and measures of accuracy then use the prediction errors on that validation set.

Naive Benchmark: The Average

The benchmark criterion in prediction is using the average outcome value (thereby ignoring all predictor information). In other words, the prediction for a new record is simply the average across the outcome values of the records in

EVALUATING PREDICTIVE PERFORMANCE 119

the training set (ȳ). This is sometimes called a naive benchmark. A good pre- dictive model should outperform the benchmark criterion in terms of predictive accuracy.

Prediction Accuracy Measures

The prediction error for record i is defined as the difference between its actual outcome value and its predicted outcome value: ei = yi − ŷi. A few popular numerical measures of predictive accuracy are:

MAE (mean absolute error/deviation) = 1 n

∑n i=1 |ei|. This gives the mag-

nitude of the average absolute error.

Mean Error = 1 n

∑n i=1 ei. This measure is similar to MAE except that

it retains the sign of the errors, so that negative errors cancel out positive errors of the same magnitude. It therefore gives an indication of whether the predictions are on average over- or underpredicting the outcome variable.

MPE (mean percentage error) = 100× 1 n

∑n i=1 ei/yi. This gives the per-

centage score of how predictions deviate from the actual values (on average), taking into account the direction of the error.

MAPE (mean absolute percentage error) = 100 × 1 n

∑n i=1 |ei/yi|. This

measure gives a percentage score of how predictions deviate (on average) from the actual values.

RMSE (root mean squared error) = √

1 n

∑n i=1 e

2 i . This is similar to the

standard error of estimate in linear regression, except that it is computed on the validation data rather than on the training data. It has the same units as the outcome variable.

Such measures can be used to compare models and to assess their degree of prediction accuracy. Note that all these measures are influenced by outliers. To check outlier influence, we can compute median-based measures (and compare to the above mean-based measures) or simply plot a histogram or boxplot of the errors. Plotting the prediction errors’ distribution is in fact very useful and can highlight more information than the metrics alone.

To illustrate the use of predictive accuracy measures and charts of prediction error distribution, consider the error metrics and charts shown in Table 5.1 and Figure 5.1. These are the result of fitting a certain predictive model to prices of used Toyota Corolla cars. The training set includes 600 cars and the validation set includes 400 cars. Results are displayed separately for the training and validation sets. We can see from the histogram and boxplot corresponding to the validation set that most errors are in the [−1000, 1000] range, with a few large positive (under-prediction) errors.

120 EVALUATING PREDICTIVE PERFORMANCE

TABLE 5.1 PREDICTION ERROR METRICS FROM A MODEL FOR TOYOTA CAR PRICES. TRAINING AND VALIDATION

code for accuracy measure

# package forecast is required to evaluate performance library(forecast)

# load file toyota.corolla.df <- read.csv("ToyotaCorolla.csv")

# randomly generate training and validation sets training <- sample(toyota.corolla.df$Id, 600) validation <- sample(setdiff(toyota.corolla.df$Id, training), 400)

# run linear regression model reg <- lm(Price~., data=toyota.corolla.df[,-c(1,2,8,11)], subset=training,

na.action=na.exclude) pred_t <- predict(reg, na.action=na.pass) pred_v <- predict(reg, newdata=toyota.corolla.df[validation,-c(1,2,8,11)],

na.action=na.pass)

## evaluate performance # training accuracy(pred_t, toyota.corolla.df[training,]$Price) # validation accuracy(pred_v, toyota.corolla.df[validation,]$Price)

Output

> # training > accuracy(pred_t, toyota.corolla.df[training,]$Price)

ME RMSE MAE MPE MAPE Test set -3.542119e-11 1012.578 784.1166 -0.8180893 7.773598

> # validation > accuracy(pred_v, toyota.corolla.df[validation,]$Price)

ME RMSE MAE MPE MAPE Test set -107.8089 1262.623 833.3621 -2.345028 8.846669

Tr ai

ni ng

−10000 −5000 0 5000 10000

0 10

20 30

40 50

V al

id at

io n

−10000 −5000 0 5000 10000

0 10

20 30

40

Training Validation

− 60

00 −

40 00

− 20

00 0

20 00

40 00

FIGURE 5.1 HISTOGRAMS AND BOXPLOTS OF TOYOTA PRICE PREDICTION ERRORS, FOR TRAINING AND VALIDATION SETS

EVALUATING PREDICTIVE PERFORMANCE 121

Comparing Training and Validation Performance

Errors that are based on the training set tell us about model fit, whereas those that are based on the validation set (called “prediction errors”) measure the model’s ability to predict new data (predictive performance). We expect training errors to be smaller than the validation errors (because the model was fitted using the training set), and the more complex the model, the greater the likelihood that it will overfit the training data (indicated by a greater difference between the training and validation errors). In an extreme case of overfitting, the training errors would be zero (perfect fit of the model to the training data), and the validation errors would be non-zero and non-negligible. For this reason, it is important to compare the error plots and metrics (RMSE, MAE, etc.) of the training and validation sets. Table 5.1 illustrates this comparison: the training set performance measures appear much lower (better) than those for the validation set. However, the charts reveal more than the metrics alone: looking at the charts in Figure 5.1, the discrepancies are most likely due to some outliers in the validation set, especially the asymmetric outliers in the validation set. The positive validation errors (under-predictions) are slightly larger than the training errors, as reflected by the medians and outliers.

Lift Chart

In some applications, the goal is to search, among a set of new records, for a subset of records that gives the highest cumulative predicted values. In such cases, a graphical way to assess predictive performance is through a lift chart. This compares the model’s predictive performance to a baseline model that has no predictors. A lift chart for a continuous response is relevant only when we are searching for a set of records that gives the highest cumulative predicted values. A lift chart is not relevant if we are interested in predicting the outcome value for each new record.

To illustrate this type of goal, called ranking, consider a car rental firm that renews its fleet regularly so that customers drive late-model cars. This entails disposing of a large quantity of used vehicles on a continuing basis. Since the firm is not primarily in the used car sales business, it tries to dispose of as much of its fleet as possible through volume sales to used car dealers. However, it is profitable to sell a limited number of cars through its own channels. Its volume deals with the used car dealers allow it flexibility to pick and choose which cars to sell in this fashion, so it would like to have a model for selecting cars for resale through its own channels. Since all cars were purchased some time ago and the deals with the used car dealers are for fixed prices (specifying a given number of cars of a certain make and model class), the cars’ costs are now irrelevant and the dealer is interested only in maximizing revenue. This is done by selecting for its

122 EVALUATING PREDICTIVE PERFORMANCE

own resale, the cars likely to generate the most revenue. The lift chart in this case gives the predicted lift for revenue.

The lift chart is based on ordering the set of records of interest (typically validation data) by their predicted value, from high to low. Then, we accumulate the actual values and plot their cumulative value on the y-axis as a function of the number of records accumulated (the x-axis value). This curve is compared to assigning a naive prediction (ȳ) to each record and accumulating these average values, which results in a diagonal line. The further away the lift curve from the diagonal benchmark line, the better the model is doing in separating records with high value outcomes from those with low value outcomes. The same information can be presented in a decile lift chart, where the ordered records are grouped into ten deciles, and for each decile, the chart presents the ratio of model lift to naive benchmark lift.

Figure 5.2 shows a lift chart and decile lift chart based on fitting a linear regression model to the Toyota data. The charts are based on the validation data of 400 cars. It can be seen that the model’s predictive performance in terms of lift is better than the baseline model, since its lift curve is higher than that of the baseline model. The lift and decile charts in Figure 5.2 would be useful in the following scenario: choosing the top 10% of the cars that gave the highest predicted sales, for example, we would gain 1.7 times the amount of revenue, compared to choosing 10% of the cars at random. This can be seen from the decile chart (Figure 5.2). This number can also be computed from the lift chart by comparing the sales for 40 random cars (the value of the baseline curve at x = 40), which is $486,871 (= the sum of the actual sales for the 400 validation set cars divided by 10) with the actual sales of the 40 cars that have the highest predicted values (the value of the lift curve at x = 40), $835,883. The ratio between these numbers is 1.7.

5.3 Judging Classifier Performance

The need for performance measures arises from the wide choice of classifiers and predictive methods. Not only do we have several different methods, but even within a single method there are usually many options that can lead to completely different results. A simple example is the choice of predictors used within a particular predictive algorithm. Before we study these various algorithms in detail and face decisions on how to set these options, we need to know how we will measure success.

A natural criterion for judging the performance of a classifier is the prob- ability of making a misclassification error. Misclassification means that the record belongs to one class but the model classifies it as a member of a different class. A classifier that makes no errors would be perfect, but we do not expect to be

JUDGING CLASSIFIER PERFORMANCE 123

0 100 200 300 400

0 10

00 00

0 30

00 00

0

Lift Chart

# cases

C um

ul at

iv e

P ric

e

10 30 50 70 90

Decile-wise lift chart

Percentile

M ea

n R

es po

ns e

0. 0

0. 5

1. 0

1. 5

FIGURE 5.2 LIFT CHART (LEFT) AND DECILE LIFT CHART (RIGHT) FOR CONTINUOUS OUTCOME VARIABLE (SALES OF TOYOTA CARS)

code for generating a lift chart and decile-wise lift chart

toyota.corolla.df <- read.csv("ToyotaCorolla.csv")

# remove missing Price data toyota.corolla.df <-

toyota.corolla.df[!is.na(toyota.corolla.df[validation,]$Price),]

# generate random Training and Validation sets training <- sample(toyota.corolla.df$Id, 600) validation <- sample(toyota.corolla.df$Id, 400)

# regression model based on all numerical predictors reg <- lm(Price~., data = toyota.corolla.df[,-c(1,2,8,11)], subset = training)

# predictions pred_v <- predict(reg, newdata = toyota.corolla.df[validation,-c(1,2,8,11)])

# load package gains, compute gains (we will use package caret for categorical y later) library(gains) gain <- gains(toyota.corolla.df[validation,]$Price[!is.na(pred_v)], pred_v[!is.na(pred_v)])

# cumulative lift chart options(scipen=999) # avoid scientific notation # we will compute the gain relative to price price <- toyota.corolla.df[validation,]$Price[!is.na(toyota.corolla.df[validation,]$Price)] plot(c(0,gain$cume.pct.of.total*sum(price))~c(0,gain$cume.obs),

xlab="# cases", ylab="Cumulative Price", main="Lift Chart", type="l")

# baseline lines(c(0,sum(price))~c(0,dim(toyota.corolla.df[validation,])[1]), col="gray", lty=2)

# Decile-wise lift chart barplot(gain$mean.resp/mean(price), names.arg = gain$depth,

xlab = "Percentile", ylab = "Mean Response", main = "Decile-wise lift chart")

124 EVALUATING PREDICTIVE PERFORMANCE

able to construct such classifiers in the real world due to “noise” and not hav- ing all the information needed to classify records precisely. Is there a minimal probability of misclassification that we should require of a classifier?

Benchmark: The Naive Rule

A very simple rule for classifying a record into one of m classes, ignoring all pre- dictor information (x1, x2, . . . , xp) that we may have, is to classify the record as a member of the majority class. In other words, “classify as belonging to the most prevalent class.” The naive rule is used mainly as a baseline or benchmark for evaluating the performance of more complicated classifiers. Clearly, a clas- sifier that uses external predictor information (on top of the class membership allocation) should outperform the naive rule. There are various performance measures based on the naive rule that measure how much better than the naive rule a certain classifier performs. One example is multiple R2, which measures the distance between the fit of the classifier to the data and the fit of the naive rule to the data (for further details, see Section 10.6).

Similar to using the sample mean (ȳ) as the naive benchmark in the numerical outcome case, the naive rule for classification relies solely on the y information and excludes any additional predictor information.

Class Separation

If the classes are well separated by the predictor information, even a small dataset will suffice in finding a good classifier, whereas if the classes are not separated at all by the predictors, even a very large dataset will not help. Figure 5.3 illustrates this for a two-class case. The top panel includes a small dataset (n = 24 records) where two predictors (income and lot size) are used for separating owners from nonowners [we thank Dean Wichern for this example, described in Johnson and Wichern (2002)]. Here, the predictor information seems useful in that it separates the two classes (owners/nonowners). The bottom panel shows a much larger dataset (n = 5000 records) where the two predictors (income and monthly average credit card spending) do not separate the two classes well in most of the higher ranges (loan acceptors/nonacceptors).

The Confusion (Classification) Matrix

In practice, most accuracy measures are derived from the confusion matrix, also called classification matrix. This matrix summarizes the correct and incorrect clas- sifications that a classifier produced for a certain dataset. Rows and columns of the confusion matrix correspond to the predicted and true (actual) classes, respectively. Table 5.2 shows an example of a classification (confusion) matrix for a two-class (0/1) problem resulting from applying a certain classifier to 3000

JUDGING CLASSIFIER PERFORMANCE 125

FIGURE 5.3 HIGH (TOP) AND LOW (BOTTOM) LEVELS OF SEPARATION BETWEEN TWO CLASSES, USING TWO PREDICTORS

TABLE 5.2 CONFUSION MATRIX BASED ON 3000 RECORDS AND TWO CLASSES

Actual Class

0 1

Predicted Class 0 2689 85

1 25 201

126 EVALUATING PREDICTIVE PERFORMANCE

records. The two diagonal cells (upper left, lower right) give the number of correct classifications, where the predicted class coincides with the actual class of the record. The off-diagonal cells give counts of misclassification. The lower left cell gives the number of class 1 members that were misclassified as 0’s (in this example, there were 85 such misclassifications). Similarly, the top-right cell gives the number of class 0 members that were misclassified as 1’s (25 such records). In R, we can obtain a confusion matrix using the function confusionMatrix() in the caret package. This function creates the cross-tabulation of actual and predicted classes. We will see an example later in this chapter.

The confusion matrix gives estimates of the true classification and misclas- sification rates. Of course, these are estimates and they can be incorrect, but if we have a large enough dataset and neither class is very rare, our estimates will be reliable. Sometimes, we may be able to use public data such as US Census data to estimate these proportions. However, in most business settings, we will not know them.

Using the Validation Data

To obtain an honest estimate of future classification error, we use the confusion matrix that is computed from the validation data. In other words, we first partition the data into training and validation sets by random selection of records. We then construct a classifier using the training data, and then apply it to the validation data. This will yield the predicted classifications for records in the validation set (see Figure 2.4 in Chapter 2). We then summarize these classifications in a confusion matrix. Although we can summarize our results in a confusion matrix for training data as well, the resulting confusion matrix is not useful for getting an honest estimate of the misclassification rate for new data due to the danger of overfitting.

In addition to examining the validation data confusion matrix to assess the classification performance on new data, we compare the training data confusion matrix to the validation data confusion matrix, in order to detect overfitting: although we expect somewhat inferior results on the validation data, a large discrepancy in training and validation performance might be indicative of over- fitting.

Accuracy Measures

Different accuracy measures can be derived from the classification matrix. Con- sider a two-class case with classes C1 and C2 (e.g., buyer/non-buyer). The schematic confusion matrix in Table 5.3 uses the notation ni,j to denote the number of records that are class Ci members and were classified as Cj members. Of course, if i ̸= j, these are counts of misclassifications. The total number of records is n = n1,1 + n1,2 + n2,1 + n2,2.

JUDGING CLASSIFIER PERFORMANCE 127

TABLE 5.3 CONFUSION MATRIX: MEANING OF EACH CELL

Actual Class

C1 C2

Predicted Class C1 n1,1 = number of C1 records n2,1 = number of C2 records classified correctly classified incorrectly as C1

C2 n1,2 = number of C1 records n2,2 = number of C2 records classified incorrectly as C2 classified correctly

A main accuracy measure is the estimated misclassification rate, also called the overall error rate. It is given by

err = n1,2 + n2,1

n ,

where n is the total number of records in the validation dataset. In the example in Table 5.2, we get err = (25 + 85)/3000 = 3.67%.

We can measure accuracy by looking at the correct classifications—the full half of the cup—instead of the misclassifications. The overall accuracy of a classifier is estimated by

accuracy = 1− err = n1,1 + n2,2 n

.

In the example, we have (201 + 2689)/3000 = 96.33%.

Propensities and Cutoff for Classification

The first step in most classification algorithms is to estimate the probability that a record belongs to each of the classes. These probabilities are also called propen- sities. Propensities are typically used either as an interim step for generating predicted class membership (classification), or for rank-ordering the records by their probability of belonging to a class of interest. Let us consider their first use in this section. The second use is discussed in Section 5.4.

If overall classification accuracy (involving all the classes) is of interest, the record can be assigned to the class with the highest probability. In many records, a single class is of special interest, so we will focus on that particular class and compare the propensity of belonging to that class to a cutoff value set by the analyst. This approach can be used with two classes or more than two classes, though it may make sense in such cases to consolidate classes so that you end up with two: the class of interest and all other classes. If the probability of belonging to the class of interest is above the cutoff, the record is assigned to that class.

128 EVALUATING PREDICTIVE PERFORMANCE

C U T O F F V A L U E S F O R T R I A G E

In some cases, it is useful to have two cutoffs, and allow a “cannot say” option for the classifier. In a two-class situation, this means that for a record, we can make one of three predictions: The record belongs to C1, or the record belongs to C2, or we cannot make a prediction because there is not enough information to pick C1 or C2 confidently. Records that the classifier cannot classify are subjected to closer scrutiny either by using expert judgment or by enriching the set of predic- tor variables by gathering additional information that is perhaps more difficult or expensive to obtain. An example is classification of documents found during legal discovery (reciprocal forced document disclosure in a legal proceeding). Under tra- ditional human-review systems, qualified legal personnel are needed to review what might be tens of thousands of documents to determine their relevance to a case. Using a classifier and a triage outcome, documents could be sorted into clearly rel- evant, clearly not relevant, and the gray area documents requiring human review. This substantially reduces the costs of discovery.

The default cutoff value in two-class classifiers is 0.5. Thus, if the probability of a record being a class C1 member is greater than 0.5, that record is classified as a C1. Any record with an estimated probability of less than 0.5 would be classified as a C2. It is possible, however, to use a cutoff that is either higher or lower than 0.5. A cutoff greater than 0.5 will end up classifying fewer records as C1’s, whereas a cutoff less than 0.5 will end up classifying more records as C1. Typically, the misclassification rate will rise in either case.

Consider the data in Table 5.4, showing the actual class for 24 records, sorted by the probability that the record is an “owner” (as estimated by a data mining algorithm). If we adopt the standard 0.5 as the cutoff, our misclassification rate is 3/24, whereas if we instead adopt a cutoff of 0.25, we classify more records as owners and the misclassification rate goes up (comprising more nonowners

TABLE 5.4 24 RECORDS WITH THEIR ACTUAL CLASS AND THE PROBABILITY (PROPENSITY) OF THEM BEING CLASS “OWNER” MEMBERS, AS ESTIMATED BY A CLASSIFIER

Actual Class Probability of Class “owner” Actual Class Probability of Class “owner”

owner 0.9959 owner 0.5055 owner 0.9875 nonowner 0.4713 owner 0.9844 nonowner 0.3371 owner 0.9804 owner 0.2179 owner 0.9481 nonowner 0.1992 owner 0.8892 nonowner 0.1494 owner 0.8476 nonowner 0.0479 nonowner 0.7628 nonowner 0.0383 owner 0.7069 nonowner 0.0248 owner 0.6807 nonowner 0.0218 owner 0.6563 nonowner 0.0161 nonowner 0.6224 nonowner 0.0031

JUDGING CLASSIFIER PERFORMANCE 129

misclassified as owners) to 5/24. Conversely, if we adopt a cutoff of 0.75, we classify fewer records as owners. The misclassification rate goes up (comprising more owners misclassified as nonowners) to 6/24. All this can be seen in the classification tables in Table 5.5.

To see the entire range of cutoff values and how the accuracy or misclassi- fication rates change as a function of the cutoff, we can plot the performance measure of interest vs. the cutoff. The results for the riding mowers example are shown in Figure 5.4. We can see that the accuracy level is pretty stable around 0.8 for cutoff values between 0.2 and 0.8.

Why would we want to use cutoff values different from 0.5 if they increase the misclassification rate? The answer is that it might be more important to classify owners properly than nonowners, and we would tolerate a greater mis- classification of the latter. Or the reverse might be true; in other words, the costs of misclassification might be asymmetric. We can adjust the cutoff value in such a case to classify more records as the high-value class, that is, accept

TABLE 5.5 CONFUSION MATRICES BASED ON CUTOFFS OF 0.5, 0.25, AND 0.75 (RIDING MOWERS EXAMPLE)

> owner.df <- read.csv("ownerExample.csv") ## cutoff = 0.5 > confusionMatrix(ifelse(owner.df$Probability>0.5, 'owner', 'nonowner'), owner.df$Class) # note: "reference" = "actual" Confusion Matrix and Statistics

Reference Prediction nonowner owner nonowner 10 1 owner 2 11

Accuracy : 0.875

## cutoff = 0.25 > confusionMatrix(ifelse(owner.df$Probability>0.25, 'owner', 'nonowner'), owner.df$Class) Confusion Matrix and Statistics

Reference Prediction nonowner owner nonowner 8 1 owner 4 11

Accuracy : 0.7916667

## cutoff = 0.75 > confusionMatrix(ifelse(owner.df$Probability>0.75, 'owner', 'nonowner'), owner.df$Class) Confusion Matrix and Statistics

Reference Prediction nonowner owner nonowner 11 5 owner 1 7

Accuracy : 0.75

130 EVALUATING PREDICTIVE PERFORMANCE

0.0 0.2 0.4 0.6 0.8 1.0

0. 0

0. 2

0. 4

0. 6

0. 8

1. 0

Cutoff Value

accuracy

overall error

FIGURE 5.4 PLOTTING ACCURACY AND OVERALL ERROR AS A FUNCTION OF THE CUTOFF VALUE (RIDING MOWERS EXAMPLE)

code for creating Figure 5.4

# create empty accuracy table accT = c()

# compute accuracy per cutoff for (cut in seq(0,1,0.1)){

cm <- confusionMatrix(1 * (df$prob > cut), df$actual) accT = c(accT, cm$overall[1])

}

# plot accuracy plot(accT ~ seq(0,1,0.1), xlab = "Cutoff Value", ylab = "", type = "l", ylim = c(0, 1)) lines(1-accT ~ seq(0,1,0.1), type = "l", lty = 2) legend("topright", c("accuracy", "overall error"), lty = c(1, 2), merge = TRUE)

more misclassifications where the misclassification cost is low. Keep in mind that we are doing so after the data mining model has already been selected—we are not changing that model. It is also possible to incorporate costs into the picture before deriving the model. These subjects are discussed in greater detail below.

JUDGING CLASSIFIER PERFORMANCE 131

Performance in Case of Unequal Importance of Classes

Suppose that it is more important to predict membership correctly in class C1 than in class C2. An example is predicting the financial status (bankrupt/solvent) of firms. It may be more important to predict correctly a firm that is going bankrupt than to predict correctly a firm that is going to remain solvent. The classifier is essentially used as a system for detecting or signaling bankruptcy. In such a case, the overall accuracy is not a good measure for evaluating the classifier. Suppose that the important class is C1. The following pair of accuracy measures are the most popular:

The sensitivity (also termed recall) of a classifier is its ability to detect the important class members correctly. This is measured by n1,1/(n1,1 + n1,2), the percentage of C1 members classified correctly.

The specificity of a classifier is its ability to rule out C2 members correctly. This is measured by n2,2/(n2,1+n2,2), the percentage of C2 members clas- sified correctly.

It can be useful to plot these measures against the cutoff value in order to find a cutoff value that balances these measures.

C O M P U T I N G R A T E S : F R O M W H O S E P O I N T O F V I E W ?

Sensitivity and specificity measure the performance of a classifier from the point of view of the “classifying agency” (e.g., a company classifying cus- tomers or a hospital classifying patients). They answer the question “how well does the classifier segregate the important class members?”. It is also possible to measure accuracy from the perspective of the entity being classified (e.g., the customer or the patient), who asks “given my predicted class, what is my chance of actually belonging to that class?”, although this question is usually less relevant in a data mining application. The terms “false discovery rate” and “false omission rate” are measures of performance from the perspective of the individual entity. If C1 is the important (positive) class, then they are defined as

The false discovery rate (FDR) is the proportion of C1 predictions that are wrong, equal to n2,1/(n1,1 + n2,1). Note that this is a ratio within the row of C1 predictions (i.e., it uses only records that were classified as C1).

The false omission rate (FOR) is the proportion of C2 predictions that are wrong, equal to n1,2/(n1,2 + n2,2). Note that this is a ratio within the row of C2 predictions (i.e., it uses only records that were classified as C2).

ROC Curve A more popular method for plotting the two measures is through ROC (Receiver Operating Characteristic) curves. Starting from the

132 EVALUATING PREDICTIVE PERFORMANCE

lower left, the ROC curve plots the pairs {sensitivity, specificity} as the cut- off value descends from 1 to 0. (A typical alternative presentation is to plot 1-specificity on the x-axis, which allows 0 to be placed on the left end of the axis, and 1 on the right.) Better performance is reflected by curves that are closer to the top-left corner. The comparison curve is the diagonal, which reflects the performance of the naive rule, using varying cutoff values (i.e., setting different thresholds on the level of majority used by the majority rule). A common met- ric to summarize an ROC curve is “area under the curve (AUC),” which ranges from 1 (perfect discrimination between classes) to 0.5 (no better than the naive rule). The ROC curve for the owner/nonowner example and its corresponding AUC are shown in Figure 5.5.

Specificity

S en

si tiv

ity

0. 0

0. 2

0. 4

0. 6

0. 8

1. 0

1.0 0.8 0.6 0.4 0.2 0.0

FIGURE 5.5 ROC CURVE FOR RIDING MOWERS EXAMPLE

code for generating ROC curve and computing AUC

library(pROC) r <- roc(df$actual, df$prob) plot.roc(r)

# compute auc auc(r)

Output

> auc(r) Area under the curve: 0.9375

JUDGING CLASSIFIER PERFORMANCE 133

Asymmetric Misclassification Costs

Implicit in our discussion of the lift curve, which measures how effective we are in identifying the members of one particular class, is the assumption that the error of misclassifying a record belonging to one class is more serious than for the other class. For example, misclassifying a household as unlikely to respond to a sales offer when it belongs to the class that would respond incurs a greater cost (the opportunity cost of the foregone sale) than the converse error. In the former case, you are missing out on a sale worth perhaps tens or hundreds of dollars. In the latter, you are incurring the costs of contacting someone who will not purchase. In such a scenario, using the misclassification rate as a criterion can be misleading.

Note that we are assuming that the cost (or benefit) of making correct classi- fications is zero. At first glance, this may seem incomplete. After all, the benefit (negative cost) of classifying a buyer correctly as a buyer would seem substantial. And in other circumstances (e.g., scoring our classification algorithm to fresh data to implement our decisions), it will be appropriate to consider the actual net dollar impact of each possible classification (or misclassification). Here, how- ever, we are attempting to assess the value of a classifier in terms of classification error, so it greatly simplifies matters if we can capture all cost/benefit information in the misclassification cells. So, instead of recording the benefit of classifying a respondent household correctly, we record the cost of failing to classify it as a respondent household. It amounts to the same thing and our goal becomes the minimization of costs, whether the costs are actual costs or missed benefits (opportunity costs).

Consider the situation where the sales offer is mailed to a random sample of people for the purpose of constructing a good classifier. Suppose that the offer is accepted by 1% of those households. For these data, if a classifier simply classifies every household as a non-responder, it will have an error rate of only 1% but it will be useless in practice. A classifier that misclassifies 2% of buying households as nonbuyers and 20% of the nonbuyers as buyers would have a higher error rate but would be better if the profit from a sale is substantially higher than the cost of sending out an offer. In these situations, if we have estimates of the cost of both types of misclassification, we can use the confusion matrix to compute the expected cost of misclassification for each record in the validation data. This enables us to compare different classifiers using overall expected costs (or profits) as the criterion.

Suppose that we are considering sending an offer to 1000 more people, where on average 1% of whom respond (1). Naively classifying everyone as a 0 has an error rate of only 1%. Using a data mining routine, suppose that we can produce these classifications:

134 EVALUATING PREDICTIVE PERFORMANCE

Actual 0 Actual 1

Predicted 0 970 2

Predicted 1 20 8

These classifications have an error rate of 100× (20 + 2)/1000 = 2.2%— higher than the naive rate.

Now suppose that the profit from a responder is $10 and the cost of sending the offer is $1. Classifying everyone as a 0 still has a misclassification rate of only 1%, but yields a profit of $0. Using the data mining routine, despite the higher misclassification rate, yields a profit of $60.

The matrix of profit is as follows (nothing is sent to the predicted 0’s so there are no costs or sales in that column):

Profit Actual 0 Actual 1

Predicted 0 0 0

Predicted 1 − $20 $80

Looked at purely in terms of costs, when everyone is classified as a 0, there are no costs of sending the offer; the only costs are the opportunity costs of failing to make sales to the ten 1’s = $100. The cost (actual costs of sending the offer, plus the opportunity costs of missed sales) of using the data mining routine to select people to send the offer to is only $48, as follows:

Costs Actual 0 Actual 1

Predicted 0 0 $20

Predicted 1 $20 $8

However, this does not improve the actual classifications themselves. A better method is to change the classification rules (and hence the misclassification rates) as discussed in the preceding section, to reflect the asymmetric costs.

A popular performance measure that includes costs is the average misclassifica- tion cost, which measures the average cost of misclassification per classified record. Denote by q1 the cost of misclassifying a class C1 record (as belonging to class C2) and by q2 the cost of misclassifying a class C2 record (as belonging to class C1). The average misclassification cost is

q1n1,2 + q2n2,1 n

.

Thus, we are looking for a classifier that minimizes this quantity. This can be computed, for instance, for different cutoff values.

JUDGING CLASSIFIER PERFORMANCE 135

It turns out that the optimal parameters are affected by the misclassification costs only through the ratio of these costs. This can be seen if we write the foregoing measure slightly differently:

q1n1,2 + q2n2,1 n

= n1,2

n1,1 + n1,2

n1,1 + n1,2 n

q1 + n2,1

n2,1 + n2,2

n2,1 + n2,2 n

q2.

Minimizing this expression is equivalent to minimizing the same expression divided by a constant. If we divide by q1, it can be seen clearly that the mini- mization depends only on q2/q1 and not on their individual values. This is very practical, because in many cases it is difficult to assess the costs associated with misclassifying a C1 member and a C2 member, but estimating the ratio is easier.

This expression is a reasonable estimate of future misclassification cost if the proportions of classes C1 and C2 in the sample data are similar to the proportions of classesC1 andC2 that are expected in the future. If instead of a random sample, we draw a sample such that one class is oversampled (as described in the next section), then the sample proportions ofC1’s andC2’s will be distorted compared to the future or population. We can then correct the average misclassification cost measure for the distorted sample proportions by incorporating estimates of the true proportions (from external data or domain knowledge), denoted by p(C1) and p(C2), into the formula:

n1,2 n1,1 + n1,2

p(C1) q1 + n2,1

n2,1 + n2,2 p(C2) q2.

Using the same logic as above, it can be shown that optimizing this quantity depends on the costs only through their ratio (q2/q1) and on the prior probabil- ities only through their ratio [p(C2)/p(C1)]. This is why software packages that incorporate costs and prior probabilities might prompt the user for ratios rather than actual costs and probabilities.

Generalization to More Than Two Classes

All the comments made above about two-class classifiers extend readily to clas- sification into more than two classes. Let us suppose that we have m classes C1, C2, . . . , Cm. The confusion matrix has m rows and m columns. The mis- classification cost associated with the diagonal cells is, of course, always zero. Incorporating prior probabilities of the various classes (where now we have m such numbers) is still done in the same manner. However, evaluating misclas- sification costs becomes much more complicated: For an m-class case we have m(m− 1) types of misclassifications. Constructing a matrix of misclassification costs thus becomes prohibitively complicated.

136 EVALUATING PREDICTIVE PERFORMANCE

5.4 Judging Ranking Performance

We now turn to the predictive goal of detecting, among a set of new records, the ones most likely to belong to a class of interest. Recall that this differs from the goal of predicting class membership for each new record.

Lift Charts for Binary Data

We already introduced lift charts in the context of a numerical outcome (Section 5.2). We now describe lift charts, also called lift curves, gains curves, or gains charts, for a binary outcome. This is a more common usage than for predicted continuous outcomes. The lift curve helps us determine how effectively we can “skim the cream” by selecting a relatively small number of records and getting a relatively large portion of the responders. The input required to construct a lift curve is a validation dataset that has been “scored” by appending to each record the propensity that it will belong to a given class.

Let’s continue with the case in which a particular class is relatively rare and of much more interest than the other class: tax cheats, debt defaulters, or respon- ders to a mailing. We would like our classification model to sift through the records and sort them according to which ones are most likely to be tax cheats, responders to the mailing, and so on. We can then make more informed deci- sions. For example, we can decide how many and which tax returns to examine if looking for tax cheats. The model will give us an estimate of the extent to which we will encounter more and more non-cheaters as we proceed through the sorted data starting with the records most likely to be tax cheats. Or we can use the sorted data to decide to which potential customers a limited-budget mailing should be targeted. In other words, we are describing the case when our goal is to obtain a rank ordering among the records according to their class membership propensities.

Sorting by Propensity To construct a lift chart, we sort the set of records by propensity, in descending order. This is the propensity to belong to the important class, say C1. Then, in each row, we compute the cumulative number ofC1 members (Actual Class =C1). For example, Table 5.6 shows the 24 records ordered in descending class “1” propensity. The right-most column accumulates the number of actual 1’s. The lift chart then plots this cumulative column against the number of records.

In R, there are multiple libraries for creating lift charts. Two useful options include the caret library, which is straightforward and easy to use, but cannot be used for a numerical outcome variable. The second option is with the gains library. Figure 5.6 shows lift curves (and R code) for these two methods. Note that the caret package lift chart uses a percentage on the y-axis.

JUDGING RANKING PERFORMANCE 137

TABLE 5.6 RECORDS SORTED BY PROPENSITY OF OWNERSHIP (HIGH TO LOW) FOR THE MOWER EXAMPLE

Obs Propensity Actual Cumulative of 1 Class Actual Class

1 0.995976726 1 1 2 0.987533139 1 2 3 0.984456382 1 3 4 0.980439587 1 4 5 0.948110638 1 5 6 0.889297203 1 6 7 0.847631864 1 7 8 0.762806287 0 7 9 0.706991915 1 8

10 0.680754087 1 9 11 0.656343749 1 10 12 0.622419543 0 10 13 0.505506928 1 11 14 0.471340450 0 11 15 0.337117362 0 11 16 0.217967810 1 12 17 0.199240432 0 12 18 0.149482655 0 12 19 0.047962588 0 12 20 0.038341401 0 12 21 0.024850999 0 12 22 0.021806029 0 12 23 0.016129906 0 12 24 0.003559986 0 12

Interpreting the lift chart What is considered good or bad performance? The ideal ranking performance would place all the 1’s at the beginning (the actual 1’s would have the highest propensities and be at the top of the table), and all the 0’s at the end. A lift chart corresponding to this ideal case would be a diagonal line with slope 1 which turns into a horizontal line (once all the 1’s were accumulated). In the example, the lift curve for the best possible classifier—a classifier that makes no errors—would overlap the existing curve at the start, continue with a slope of 1 until it reached all the 12 1’s, then continue horizontally to the right.

In contrast, a useless model would be one that randomly assigns propensities (shuffling the 1’s and 0’s randomly in the Actual Class column). Such behavior would increase the cumulative number of 1’s, on average, by #1

′s n

in each row. And in fact, this is the diagonal line joining the points (0,0) to (24,12) seen in Figure 5.6. This serves as a reference line. For any given number of records (the x-axis value), it represents the expected number of 1 classifications if we did not

138 EVALUATING PREDICTIVE PERFORMANCE

% Samples Tested

% S

am pl

es F

ou nd

0

20

40

60

80

100

0 20 40 60 80 100 0 5 10 15 20

0 2

4 6

8 10

12

# cases

C um

ul at

iv e

FIGURE 5.6 LIFT CHART FOR THE MOWER EXAMPLE USING CARET PACKAGE (TOP) AND GAINS PACKAGE (BOTTOM)

code for creating a lift chart: two options

# first option with 'caret' library: library(caret) lift.example <- lift(relevel(as.factor(actual), ref="1") ~ prob, data = df) xyplot(lift.example, plot = "gain")

# Second option with 'gains' library: library(gains) df <- read.csv("liftExample.csv") gain <- gains(df$actual, df$prob, groups=dim(df)[1]) plot(c(0, gain$cume.pct.of.total*sum(df$actual)) ~ c(0, gain$cume.obs),

xlab = "# cases", ylab = "Cumulative", type="l") lines(c(0,sum(df$actual))~c(0,dim(df)[1]), col="gray", lty=2)

have a model but simply selected records at random. It provides a benchmark against which we can evaluate the ranking performance of the model. In this example, although our model is not perfect, it seems to perform much better than the random benchmark.

How do we read a lift chart? For a given number of records (x-axis), the lift curve value on the y-axis tells us how much better we are doing compared to random assignment. For example, looking at Figure 5.6, if we use our model to choose the top 10 records, the lift curve tells us that we would be right for about nine of them (or 18% using the caret package lift curve). If we simply select 10 records at random, we expect to be right for 10× 12/24 = 5 records. The model gives us a “lift” in detecting class 1 members of 9/5 = 1.8. The lift will vary with the number of records we choose to act on. A good classifier will give us a high lift when we act on only a few records. As we include more records, the lift will decrease.

Decile Lift Charts

The information from the lift chart can be portrayed as a decile chart, as shown in Figure 5.7, which is widely used in direct marketing predictive modeling. The

JUDGING RANKING PERFORMANCE 139

8 17 29 38 50 58 67 79 88 100

Decile−wise lift chart

Percentile

M ea

n R

es po

ns e

0. 0

0. 5

1. 0

1. 5

2. 0

FIGURE 5.7 DECILE LIFT CHART

code for creating a decile lift chart

# use gains() to compute deciles. # when using the caret package, deciles must be computed manually.

gain <- gains(df$actual, df$prob,) barplot(gain$mean.resp / mean(df$actual), names.arg = gain$depth, xlab = "Percentile",

ylab = "Mean Response", main = "Decile-wise lift chart")

decile chart aggregates all the lift information into 10 buckets. The dots show, on the y-axis, the factor by which our model outperforms a random assignment of 0’s and 1’s, taking one decile at a time. Reading the bar on the left, we see that taking 8% of the records that are ranked by the model as “the most probable 1’s” (having the highest propensities) yields twice as many 1’s as would a random selection of 8% of the records. In this example, the decile chart indicates that we can even use the model to select the top 29% records with the highest propensities and still perform twice as well as random.

Beyond Two Classes

A lift chart cannot be used with a multiclass classifier unless a single “important class” is defined and the classifications are reduced to “important” and “unim- portant” classes.

Lift Charts Incorporating Costs and Benefits

When the benefits and costs of correct and incorrect classification are known or can be estimated, the lift chart is still a useful presentation and decision tool. As

140 EVALUATING PREDICTIVE PERFORMANCE

before, we need a classifier that assigns to each record a propensity that it belongs to a particular class. The procedure is then as follows:

1. Sort the records in descending order of predicted probability of success (where success = belonging to the class of interest).

2. For each record, record the cost (benefit) associated with the actual out- come.

3. For the highest propensity (i.e., first) record, its x-axis value is 1 and its y-axis value is its cost or benefit (computed in Step 2) on the lift curve.

4. For the next record, again calculate the cost (benefit) associated with the actual outcome. Add this to the cost (benefit) for the previous record. This sum is the y-axis coordinate of the second point on the lift curve. Its x-axis value is 2.

5. Repeat Step 4 until all records have been examined. Connect all the points, and this is the lift curve.

6. The reference line is a straight line from the origin to the point y = total net benefit and x = n(n = number of records).

Note: It is entirely possible for a reference line that incorporates costs and ben- efits to have a negative slope if the net value for the entire dataset is negative. For example, if the cost of mailing to a person is $0.65, the value of a responder is $25, and the overall response rate is 2%, the expected net value of mailing to a list of 10,000 is (0.02× $25× 10, 000)− ($0.65× 10, 000) = $5000− $6500 = −$1500. Hence, the y-value at the far right of the lift curve (x = 10,000) is −1500, and the slope of the reference line from the origin will be negative. The optimal point will be where the lift curve is at a maximum (i.e., mailing to about 3000 people) in Figure 5.8.

Lift as a Function of Cutoff

We could also plot the lift as a function of the cutoff value. The only difference is the scale on the x-axis. When the goal is to select the top records based on a certain budget, the lift vs. number of records is preferable. In contrast, when the goal is to find a cutoff that distinguishes well between the two classes, the lift vs. cutoff value is more useful.

5.5 Oversampling

As we saw briefly in Chapter 2, when classes are present in very unequal pro- portions, simple random sampling may produce too few of the rare class to yield useful information about what distinguishes them from the dominant class. In such cases, stratified sampling is often used to oversample the records from the

OVERSAMPLING 141

FIGURE 5.8 LIFT CURVE INCORPORATING COSTS

rarer class and improve the performance of classifiers. It is often the case that the rarer events are the more interesting or important ones: responders to a mailing, those who commit fraud, defaulters on debt, and the like. This same stratified sampling procedure is sometimes called weighted sampling or undersampling, the latter referring to the fact that the more plentiful class is undersampled, relative to the rare class. We shall stick to the term oversampling.

In all discussions of oversampling, we assume the common situation in which there are two classes, one of much greater interest than the other. Data with more than two classes do not lend themselves to this procedure.

Consider the data in Figure 5.9, where × represents non-responders, and ◦, responders. The two axes correspond to two predictors. The dashed vertical line does the best job of classification under the assumption of equal costs: It results in just one misclassification (one ◦ is misclassified as an ×). If we incorporate more realistic misclassification costs—let’s say that failing to catch a ◦ is five times as costly as failing to catch an ×—the costs of misclassification jump to 5. In such a case, a horizontal line as shown in Figure 5.10, does a better job: It results in misclassification costs of just 2.

Oversampling is one way of incorporating these costs into the training pro- cess. In Figure 5.11, we can see that classification algorithms would automatically determine the appropriate classification line if four additional ◦’s were present at each existing ◦. We can achieve appropriate results either by taking five times as many ◦’s as we would get from simple random sampling (by sampling with replacement if necessary), or by replicating the existing ◦’s fourfold.

142 EVALUATING PREDICTIVE PERFORMANCE

FIGURE 5.9 CLASSIFICATION ASSUMING EQUAL COSTS OF MISCLASSIFICATION

FIGURE 5.10 CLASSIFICATION ASSUMING UNEQUAL COSTS OF MISCLASSIFICATION

OVERSAMPLING 143

FIGURE 5.11 CLASSIFICATION USING OVERSAMPLING TO ACCOUNT FOR UNEQUAL COSTS

Oversampling without replacement in accord with the ratio of costs (the first option above) is the optimal solution, but may not always be practical. There may not be an adequate number of responders to assure that there will be enough of them to fit a model if they constitute only a small proportion of the total. Also, it is often the case that our interest in discovering responders is known to be much greater than our interest in discovering non-responders, but the exact ratio of costs is difficult to determine. When faced with very low response rates in a classification problem, practitioners often sample equal numbers of responders and non-responders as a relatively effective and convenient approach. Whatever approach is used, when it comes time to assess and predict model performance, we will need to adjust for the oversampling in one of two ways:

1. score the model to a validation set that has been selected without over- sampling (i.e., via simple random sampling), or

2. score the model to an oversampled validation set, and reweight the results to remove the effects of oversampling.

The first method is more straightforward and easier to implement. We describe how to oversample and how to evaluate performance for each of the two methods.

144 EVALUATING PREDICTIVE PERFORMANCE

When classifying data with very low response rates, practitioners typically:

• train models on data that are 50% responder, 50% non-responder • validate the models with an unweighted (simple random) sample from the original

data

Oversampling the Training Set

How is weighted sampling done? When responders are sufficiently scarce that you will want to use all of them, one common procedure is as follows:

1. First, the response and non-response data are separated into two distinct sets, or strata.

2. Records are then randomly selected for the training set from each stratum. Typically, one might select half the (scarce) responders for the training set, then an equal number of non-responders.

3. The remaining responders are put in the validation set.

4. Non-responders are randomly selected for the validation set in sufficient numbers to maintain the original ratio of responders to non-responders.

5. If a test set is required, it can be taken randomly from the validation set.

Evaluating Model Performance Using a Non-oversampled Validation Set

Although the oversampled data can be used to train models, they are often not suitable for evaluating model performance, because the number of responders will (of course) be exaggerated. The most straightforward way of gaining an unbiased estimate of model performance is to apply the model to regular data (i.e., data not oversampled). In short, train the model on oversampled data, but validate it with regular data.

Evaluating Model Performance if Only Oversampled Validation Set Exists

In some cases, very low response rates may make it more practical to use over- sampled data not only for the training data, but also for the validation data. This might happen, for example, if an analyst is given a dataset for exploration and prototyping that is already oversampled to boost the proportion with the rare response of interest (perhaps because it is more convenient to transfer and work with a smaller dataset). In such cases, it is still possible to assess how well the model will do with real data, but this requires the oversampled validation set to be reweighted, in order to restore the class of records that were underrepre- sented in the sampling process. This adjustment should be made to the confusion matrix and to the lift chart in order to derive good accuracy measures. These adjustments are described next.

OVERSAMPLING 145

I. Adjusting the Confusion Matrix for Oversampling Suppose the response rate in the data as a whole is 2%, and that the data were oversampled, yielding a sample in which the response rate is 25 times higher (50% responders). The relationship is as follows:

Responders: 2% of the whole data; 50% of the sample

Non-responders: 98% of the whole data, 50% of the sample

Each responder in the whole data is worth 25 responders in the sample (50/2). Each non-responder in the whole data is worth 0.5102 non-responders in the sample (50/98). We call these values oversampling weights.

Assume that the validation confusion matrix looks like this:

CONFUSION MATRIX, OVERSAMPLED DATA (VALIDATION)

Actual 0 Actual 1 Total

Predicted 0 390 80 470

Predicted 1 110 420 530

Total 500 500 1000

At this point, the misclassification rate appears to be (80 + 110)/1000 = 19%, and the model ends up classifying 53% of the records as 1’s. However, this reflects the performance on a sample where 50% are responders.

To estimate predictive performance when this model is used to score the original population (with 2% responders), we need to undo the effects of the oversampling. The actual number of responders must be divided by 25, and the actual number of non-responders divided by 0.5102.

The revised confusion matrix is as follows:

CONFUSION MATRIX, REWEIGHTED

Actual 0 Actual 1 Total

Predicted 0 390/0.5102 = 764.4 80/25 = 3.2

Predicted 1 110/0.5102 = 215.6 420/25 = 16.8

Total 980 20 1000

The adjusted misclassification rate is (3.2 + 215.6)/1000 = 21.9%. The model ends up classifying (215.6+16.8)/1000 = 23.24% of the records as 1’s, when we assume 2% responders.

146 EVALUATING PREDICTIVE PERFORMANCE

II. Adjusting the Lift Curve for Oversampling The lift curve is likely to be a more useful measure in low-response situations, where our interest lies not so much in classifying all the records correctly as in finding a model that guides us toward those records most likely to contain the response of interest (under the assumption that scarce resources preclude examining or contacting all the records). Typically, our interest in such a case is in maximizing value or minimizing cost, so we will show the adjustment process incorporating the benefit/cost element. The following procedure can be used:

1. Sort the validation records in order of the predicted probability of success (where success = belonging to the class of interest).

2. For each record, record the cost (benefit) associated with the actual out- come.

3. Divide that value by the oversampling rate. For example, if responders are overweighted by a factor of 25, divide by 25.

4. For the highest probability (i.e., first) record, the value above is the y- coordinate of the first point on the lift chart. The x-coordinate is index number 1.

5. For the next record, again calculate the adjusted value associated with the actual outcome. Add this to the adjusted cost (benefit) for the previous record. This sum is the y-coordinate of the second point on the lift curve. The x-coordinate is index number 2.

6. Repeat Step 5 until all records have been examined. Connect all the points, and this is the lift curve.

7. The reference line is a straight line from the origin to the point y = total net benefit and x = n (n = number of records).

PROBLEMS 147

PROBLEMS

5.1 A data mining routine has been applied to a transaction dataset and has classified 88 records as fraudulent (30 correctly so) and 952 as non-fraudulent (920 correctly so). Construct the confusion matrix and calculate the overall error rate.

5.2 Suppose that this routine has an adjustable cutoff (threshold) mechanism by which you can alter the proportion of records classified as fraudulent. Describe how moving the cutoff up or down would affect

a. the classification error rate for records that are truly fraudulent

b. the classification error rate for records that are truly nonfraudulent

5.3 FiscalNote is a startup founded by a Washington, DC entrepreneur and funded by a Singapore sovereign wealth fund, the Winklevoss twins of Facebook fame, and oth- ers. It uses machine learning and data mining techniques to predict for its clients whether legislation in the US Congress and in US state legislatures will pass or not. The company reports 94% accuracy. (Washington Post, November 21, 2014, “Capital Business”)

Considering just bills introduced in the US Congress, do a bit of internet research to learn about numbers of bills introduced and passage rates. Identify the possible types of misclassifications, and comment on the use of overall accuracy as a metric. Include a discussion of other possible metrics and the potential role of propensities.

5.4 Consider Figure 5.12, the decile-wise lift chart for the transaction data model, applied to new data.

1 2 3 4 5 6 7 8 9 10

Percentile

M ea

n R

es po

ns e

0 1

2 3

4 5

6

FIGURE 5.12 DECILE-WISE LIFT CHART FOR TRANSACTION DATA

a. Interpret the meaning of the first and second bars from the left.

b. Explain how you might use this information in practice.

c. Another analyst comments that you could improve the accuracy of the model by classifying everything as nonfraudulent. If you do that, what is the error rate?

d. Comment on the usefulness, in this situation, of these two metrics of model per- formance (error rate and lift).

5.5 A large number of insurance records are to be examined to develop a model for pre- dicting fraudulent claims. Of the claims in the historical database, 1% were judged

148 EVALUATING PREDICTIVE PERFORMANCE

to be fraudulent. A sample is taken to develop a model, and oversampling is used to provide a balanced sample in light of the very low response rate. When applied to this sample (n = 800), the model ends up correctly classifying 310 frauds, and 270 nonfrauds. It missed 90 frauds, and classified 130 records incorrectly as frauds when they were not.

a. Produce the confusion matrix for the sample as it stands.

b. Find the adjusted misclassification rate (adjusting for the oversampling).

c. What percentage of new records would you expect to be classified as fraudulent?

5.6 A firm that sells software services has been piloting a new product and has records of 500 customers who have either bought the services or decided not to. The target value is the estimated profit from each sale (excluding sales costs). The global mean is $2128. However, the cost of the sales effort is not cheap—the company figures it comes to $2500 for each of the 500 customers (whether they buy or not). The firm developed a predictive model in hopes of being able to identify the top spenders in the future. The lift and decile charts for the validation set are shown in Figure 5.13.

0 50 100 150 200

0 10

00 00

30 00

00 50

00 00

# cases

C um

ul at

iv e

10 20 30 40 50 60 70 80 90 100

Decile-wise lift chart

Percentile

M ea

n R

es po

ns e

0. 0

0. 5

1. 0

1. 5

2. 0

FIGURE 5.13 LIFT AND DECILE-WISE LIFT CHARTS FOR SOFTWARE SERVICES PRODUCT SALES

a. If the company begins working with a new set of 1000 leads to sell the same services, similar to the 500 in the pilot study, without any use of predictive modeling to target sales efforts, what is the estimated profit?

b. If the firm wants the average profit on each sale to at least double the sales effort cost, and applies an appropriate cutoff with this predictive model to a new set of 1000 leads, how far down the new list of 1000 should it proceed (how many deciles)?

c. Still considering the new list of 1000 leads, if the company applies this predictive model with a lower cutoff of $2500, how far should it proceed down the ranked leads, in terms of deciles?

PROBLEMS 149

d. Why use this two-stage process for predicting sales—why not simply develop a model for predicting profit for the 1000 new leads?

5.7 Table 5.7 shows a small set of predictive model validation results for a classification model, with both actual values and propensities.

a. Calculate error rates, sensitivity, and specificity using cutoffs of 0.25, 0.5, and 0.75.

b. Create a decile-wise lift chart in R.

TABLE 5.7 PROPENSITIES AND ACTUAL CLASS MEMBERSHIP FOR VALIDATION DATA

Propensity of 1 Actual

0.03 0 0.52 0 0.38 0 0.82 1 0.33 0 0.42 0 0.55 1 0.59 0 0.09 0 0.21 0 0.43 0 0.04 0 0.08 0 0.13 0 0.01 0 0.79 1 0.42 0 0.29 0 0.08 0 0.02 0

Part IV

Prediction and Classification Methods

CHAPTER 6

Multiple Linear Regression

In this chapter, we introduce linear regression models for the purpose of pre- diction. We discuss the differences between fitting and using regression models for the purpose of inference (as in classical statistics) and for prediction. A pre- dictive goal calls for evaluating model performance on a validation set, and for using predictive metrics. We then raise the challenges of using many predictors and describe variable selection algorithms that are often implemented in linear regression procedures.

6.1 Introduction

The most popular model for making predictions is the multiple linear regres- sion model encountered in most introductory statistics courses and textbooks. This model is used to fit a relationship between a numerical outcome variable Y (also called the response, target, or dependent variable) and a set of predictors X1, X2, . . . , Xp (also referred to as independent variables, input variables, regres- sors, or covariates). The assumption is that the following function approximates the relationship between the predictors and outcome variable:

Y = β0 + β1x1 + β2x2 + · · ·+ βpxp + ϵ, (6.1)

where β0, . . . , βp are coefficients and ϵ is the noise or unexplained part. Data are then used to estimate the coefficients and to quantify the noise. In predictive modeling, the data are also used to evaluate model performance.

Regression modeling means not only estimating the coefficients but also choosing which predictors to include and in what form. For example, a numer- ical predictor can be included as is, or in logarithmic form [log(X)], or in a

153

154 MULTIPLE LINEAR REGRESSION

binned form (e.g., age group). Choosing the right form depends on domain knowledge, data availability, and needed predictive power.

Multiple linear regression is applicable to numerous predictive modeling sit- uations. Examples are predicting customer activity on credit cards from their demographics and historical activity patterns, predicting expenditures on vaca- tion travel based on historical frequent flyer data, predicting staffing require- ments at help desks based on historical data and product and sales information, predicting sales from cross-selling of products from historical information, and predicting the impact of discounts on sales in retail outlets.

6.2 Explanatory vs. Predictive Modeling

Before introducing the use of linear regression for prediction, we must clarify an important distinction that often escapes those with earlier familiarity with linear regression from courses in statistics. In particular, the two popular but different objectives behind fitting a regression model are:

1. Explaining or quantifying the average effect of inputs on an outcome (explanatory or descriptive task, respectively)

2. Predicting the outcome value for new records, given their input values (predictive task)

The classical statistical approach is focused on the first objective. In that scenario, the data are treated as a random sample from a larger population of interest. The regression model estimated from this sample is an attempt to capture the average relationship in the larger population. This model is then used in decision- making to generate statements such as “a unit increase in service speed (X1) is associated with an average increase of 5 points in customer satisfaction (Y ), all other factors (X2, X3, . . . , Xp) being equal.” If X1 is known to cause Y , then such a statement indicates actionable policy changes—this is called explanatory modeling. When the causal structure is unknown, then this model quantifies the degree of association between the inputs and outcome variable, and the approach is called descriptive modeling.

In predictive analytics, however, the focus is typically on the second goal: predicting new individual records. Here we are not interested in the coefficients themselves, nor in the “average record,” but rather in the predictions that this model can generate for new records. In this scenario, the model is used for micro-decision-making at the record level. In our previous example, we would use the regression model to predict customer satisfaction for each new customer of interest.

EXPLANATORY VS. PREDICTIVE MODELING 155

Both explanatory and predictive modeling involve using a dataset to fit a model (i.e., to estimate coefficients), checking model validity, assessing its per- formance, and comparing to other models. However, the modeling steps and performance assessment differ in the two cases, usually leading to different final models. Therefore, the choice of model is closely tied to whether the goal is explanatory or predictive.

In explanatory and descriptive modeling, where the focus is on modeling the average record, we try to fit the best model to the data in an attempt to learn about the underlying relationship in the population. In contrast, in predictive modeling (data mining), the goal is to find a regression model that best predicts new individual records. A regression model that fits the existing data too well is not likely to perform well with new data. Hence, we look for a model that has the highest predictive power by evaluating it on a holdout set and using predictive metrics (see Chapter 5).

Let us summarize the main differences in using a linear regression in the two scenarios:

1. A good explanatory model is one that fits the data closely, whereas a good predictive model is one that predicts new records accurately. Choices of input variables and their form can therefore differ.

2. In explanatory models, the entire dataset is used for estimating the best- fit model, to maximize the amount of information that we have about the hypothesized relationship in the population. When the goal is to predict outcomes of new individual records, the data are typically split into a training set and a validation set. The training set is used to estimate the model, and the validation or holdout set is used to assess this model’s predictive performance on new, unobserved data.

3. Performance measures for explanatory models measure how close the data fit the model (how well the model approximates the data) and how strong the average relationship is, whereas in predictive models perfor- mance is measured by predictive accuracy (how well the model predicts new individual records).

4. In explanatory models the focus is on the coefficients (β), whereas in predictive models the focus is on the predictions (ŷ).

For these reasons, it is extremely important to know the goal of the analysis before beginning the modeling process. A good predictive model can have a looser fit to the data on which it is based, and a good explanatory model can have low prediction accuracy. In the remainder of this chapter, we focus on predictive models because these are more popular in data mining and because most statistics textbooks focus on explanatory modeling.

156 MULTIPLE LINEAR REGRESSION

6.3 Estimating the Regression Equation and Prediction

Once we determine the predictors to include and their form, we estimate the coefficients of the regression formula from the data using a method called ordinary least squares (OLS). This method finds values β̂0, β̂1, β̂2, . . . , β̂p that minimize the sum of squared deviations between the actual outcome values (Y ) and their predicted values based on that model (Ŷ ).

To predict the value of the outcome variable for a record with predictor values x1, x2, . . . , xp, we use the equation

Ŷ = β̂0 + β̂1x1 + β̂2x2 + · · ·+ β̂pxp. (6.2)

Predictions based on this equation are the best predictions possible in the sense that they will be unbiased (equal to the true values on average) and will have the smallest mean squared error compared to any unbiased estimates if we make the following assumptions:

1. The noise ϵ (or equivalently, Y ) follows a normal distribution.

2. The choice of predictors and their form is correct (linearity).

3. The records are independent of each other.

4. The variability in the outcome values for a given set of predictors is the same regardless of the values of the predictors (homoskedasticity).

An important and interesting fact for the predictive goal is that even if we drop the first assumption and allow the noise to follow an arbitrary distribution, these estimates are very good for prediction, in the sense that among all linear models, as defined by equation (6.1), the model using the least squares estimates, β̂0, β̂1, β̂2, . . . , β̂p, will have the smallest mean squared errors. The assumption of a normal dis- tribution is required in explanatory modeling, where it is used for constructing confidence intervals and statistical tests for the model parameters.

Even if the other assumptions are violated, it is still possible that the resulting predictions are sufficiently accurate and precise for the purpose they are intended for. The key is to evaluate predictive performance of the model, which is the main priority. Satisfying assumptions is of secondary interest and residual analysis can give clues to potential improved models to examine.

Example: Predicting the Price of Used Toyota Corolla Cars

A large Toyota car dealership offers purchasers of new Toyota cars the option to buy their used car as part of a trade-in. In particular, a new promotion promises

ESTIMATING THE REGRESSION EQUATION AND PREDICTION 157

TABLE 6.1 VARIABLES IN THE TOYOTA COROLLA EXAMPLE

Variable Description

Price Offer price in Euros

Age Age in months as of August 2004

Kilometers Accumulated kilometers on odometer

Fuel Type Fuel type (Petrol, Diesel, CNG)

HP Horsepower

Metallic Metallic color? (Yes = 1, No = 0)

Automatic Automatic (Yes = 1, No = 0)

CC Cylinder volume in cubic centimeters

Doors Number of doors

QuartTax Quarterly road tax in Euros

Weight Weight in kilograms

to pay high prices for used Toyota Corolla cars for purchasers of a new car. The dealer then sells the used cars for a small profit. To ensure a reasonable profit, the dealer needs to be able to predict the price that the dealership will get for the used cars. For that reason, data were collected on all previous sales of used Toyota Corollas at the dealership. The data include the sales price and other information on the car, such as its age, mileage, fuel type, and engine size. A description of each of these variables is given in Table 6.1. A sam- ple of this dataset is shown in Table 6.2. The total number of records in the dataset is 1000 cars (we use the first 1000 cars from the dataset ToyotoCorolla.csv). After partitioning the data into training (60%) and validation (40%) sets, we fit a multiple linear regression model between price (the outcome variable) and the other variables (as predictors) using only the training set. Table 6.3 shows the estimated coefficients. Notice that the Fuel Type predictor has three cate- gories (Petrol, Diesel, and CNG). We therefore have two dummy variables in the model: Fuel_TypePetrol (0/1) and Fuel_TypeDiesel (0/1); the third, for CNG (0/1), is redundant given the information on the first two dummies. Including the redundant dummy would cause the regression to fail, since the redundant dummy will be a perfect linear combination of the other two; R’s “lm” routine handles this issue automatically.

The regression coefficients are then used to predict prices of individual used Toyota Corolla cars based on their age, mileage, and so on. Table 6.4 shows a sample of predicted prices for 20 cars in the validation set, using the estimated model. It gives the predictions and their errors (relative to the actual prices) for these 20 cars. Below the predictions, we have overall measures of predictive

158 MULTIPLE LINEAR REGRESSION

TABLE 6.2 PRICES AND ATTRIBUTES FOR USED TOYOTA COROLLA CARS (SELECTED ROWS AND COLUMNS ONLY)

Fuel Auto- Quart

Price Age Kilometers Type HP Metallic matic CC Doors Tax Weight

13500 23 46986 Diesel 90 1 0 2000 3 210 1165

13750 23 72937 Diesel 90 1 0 2000 3 210 1165

13950 24 41711 Diesel 90 1 0 2000 3 210 1165

14950 26 48000 Diesel 90 0 0 2000 3 210 1165

13750 30 38500 Diesel 90 0 0 2000 3 210 1170

12950 32 61000 Diesel 90 0 0 2000 3 210 1170

16900 27 94612 Diesel 90 1 0 2000 3 210 1245

18600 30 75889 Diesel 90 1 0 2000 3 210 1245

21500 27 19700 Petrol 192 0 0 1800 3 100 1185

12950 23 71138 Diesel 69 0 0 1900 3 185 1105

20950 25 31461 Petrol 192 0 0 1800 3 100 1185

19950 22 43610 Petrol 192 0 0 1800 3 100 1185

19600 25 32189 Petrol 192 0 0 1800 3 100 1185

21500 31 23000 Petrol 192 1 0 1800 3 100 1185

22500 32 34131 Petrol 192 1 0 1800 3 100 1185

22000 28 18739 Petrol 192 0 0 1800 3 100 1185

22750 30 34000 Petrol 192 1 0 1800 3 100 1185

17950 24 21716 Petrol 110 1 0 1600 3 85 1105

16750 24 25563 Petrol 110 0 0 1600 3 19 1065

16950 30 64359 Petrol 110 1 0 1600 3 85 1105

15950 30 67660 Petrol 110 1 0 1600 3 85 1105

16950 29 43905 Petrol 110 0 1 1600 3 100 1170

15950 28 56349 Petrol 110 1 0 1600 3 85 1120

16950 28 32220 Petrol 110 1 0 1600 3 85 1120

16250 29 25813 Petrol 110 1 0 1600 3 85 1120

15950 25 28450 Petrol 110 1 0 1600 3 85 1120

17495 27 34545 Petrol 110 1 0 1600 3 85 1120

15750 29 41415 Petrol 110 1 0 1600 3 85 1120

11950 39 98823 CNG 110 1 0 1600 5 197 1119

accuracy. Note that the mean error (ME) is $ − 40 and RMSE = $1321. A histogram of the residuals (Figure 6.1) shows that most of the errors are between ±$2000. This error magnitude might be small relative to the car price, but should be taken into account when considering the profit. Another observation of interest is the large positive residuals (under-predictions), which may or may not be a concern, depending on the application. Measures such as the mean error, and error percentiles are used to assess the predictive performance of a model and to compare models.

ESTIMATING THE REGRESSION EQUATION AND PREDICTION 159

TABLE 6.3 LINEAR REGRESSION MODEL OF PRICE VS. CAR ATTRIBUTES

code for fitting a regression model

car.df <- read.csv("ToyotaCorolla.csv") # use first 1000 rows of data car.df <- car.df[1:1000, ] # select variables for regression selected.var <- c(3, 4, 7, 8, 9, 10, 12, 13, 14, 17, 18)

# partition data set.seed(1) # set seed for reproducing the partition train.index <- sample(c(1:1000), 600) train.df <- car.df[train.index, selected.var] valid.df <- car.df[-train.index, selected.var]

# use lm() to run a linear regression of Price on all 11 predictors in the # training set. # use . after ~ to include all the remaining columns in train.df as predictors. car.lm <- lm(Price ~ ., data = train.df) # use options() to ensure numbers are not displayed in scientific notation. options(scipen = 999) summary(car.lm)

Partial Output

> summary(car.lm)

Call: lm(formula = Price ~ ., data = train.df)

Residuals: Min 1Q Median 3Q Max

-8212.5 -839.2 -14.3 831.5 7270.7

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) -1774.877829 1643.744823 -1.080 0.2807 Age_08_04 -135.430875 4.875906 -27.776 < 0.0000000000000002 *** KM -0.019003 0.002341 -8.116 0.00000000000000283 *** Fuel_TypeDiesel 1208.339159 534.431400 2.261 0.0241 * Fuel_TypePetrol 2425.876714 520.587979 4.660 0.00000391697679667 *** HP 38.985537 5.587183 6.978 0.00000000000811621 *** Met_Color 84.792715 126.883452 0.668 0.5042 Automatic 306.684154 289.433138 1.060 0.2898 CC 0.031966 0.099075 0.323 0.7471 Doors -44.157742 64.056530 -0.689 0.4909 Quarterly_Tax 16.677343 2.602668 6.408 0.00000000030287017 *** Weight 12.667487 1.536587 8.244 0.00000000000000109 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1406 on 588 degrees of freedom Multiple R-squared: 0.8567, Adjusted R-squared: 0.854 F-statistic: 319.6 on 11 and 588 DF, p-value: < 0.00000000000000022

160 MULTIPLE LINEAR REGRESSION

TABLE 6.4 PREDICTED PRICES (AND ERRORS) FOR 20 CARS IN VALIDATION SET AND SUMMARY PREDICTIVE MEASURES FOR ENTIRE VALIDATION SET (CALLED TEST SET IN R)

code for prediction and measuring accuracy

library(forecast) # use predict() to make predictions on a new set. car.lm.pred <- predict(car.lm, valid.df) options(scipen=999, digits = 0) some.residuals <- valid.df$Price[1:20] - car.lm.pred[1:20] data.frame("Predicted" = car.lm.pred[1:20], "Actual" = valid.df$Price[1:20],

"Residual" = some.residuals)

options(scipen=999, digits = 3) # use accuracy() to compute common accuracy measures. accuracy(car.lm.pred, valid.df$Price)

Output

> data.frame("Predicted" = car.lm.pred[1:20], + "Actual" = valid.df$Price[1:20], "Residual" = some.residuals)

Predicted Actual Residual 3 17175 13950 -3225 6 15704 12950 -2754 8 16727 18600 1873 9 20709 21500 791 10 14668 12950 -1718 11 20756 20950 194 13 20743 19600 -1143 16 20592 22000 1408 17 20116 22750 2634 18 16695 17950 1255 19 14930 16750 1820 20 15072 16950 1878 25 16130 16250 120 26 16622 15950 -672 27 16235 17495 1260 29 15832 16950 1118 33 15564 15950 386 34 15639 14950 -689 37 15836 15950 114 38 16477 14950 -1527

> accuracy(car.lm.pred, valid.df$Price) ME RMSE MAE MPE MAPE

Test set -40.1 1321 1012 -1.72 9.01

VARIABLE SELECTION IN LINEAR REGRESSION 161

code for plotting histogram of validation errors

library(forecast) car.lm.pred <- predict(car.lm, valid.df) all.residuals <- valid.df$Price - car.lm.pred length(all.residuals[which(all.residuals > -1406 & all.residuals < 1406)])/400 hist(all.residuals, breaks = 25, xlab = "Residuals", main = "")

Residuals

F re

qu en

cy

−4000 −2000 0 2000 4000 6000

0 10

20 30

40 50

60 70

FIGURE 6.1 HISTOGRAM OF MODEL ERRORS (BASED ON VALIDATION SET)

6.4 Variable Selection in Linear Regression

Reducing the Number of Predictors

A frequent problem in data mining is that of using a regression equation to predict the value of a dependent variable when we have many variables available to choose as predictors in our model. Given the high speed of modern algorithms for multiple linear regression calculations, it is tempting in such a situation to take a kitchen-sink approach: Why bother to select a subset? Just use all the variables in the model.

Another consideration favoring the inclusions of numerous variables is the hope that a previously hidden relationship will emerge. For example, a company found that customers who had purchased anti-scuff protectors for chair and table

162 MULTIPLE LINEAR REGRESSION

legs had lower credit risks. However, there are several reasons for exercising caution before throwing all possible variables into a model.

• It may be expensive or not feasible to collect a full complement of pre- dictors for future predictions.

• We may be able to measure fewer predictors more accurately (e.g., in surveys).

• The more predictors, the higher the chance of missing values in the data. If we delete or impute records with missing values, multiple predictors will lead to a higher rate of record deletion or imputation.

• Parsimony is an important property of good models. We obtain more insight into the influence of predictors in models with few parameters.

• Estimates of regression coefficients are likely to be unstable, due to mul- ticollinearity in models with many variables. (Multicollinearity is the pres- ence of two or more predictors sharing the same linear relationship with the outcome variable.) Regression coefficients are more stable for parsi- monious models. One very rough rule of thumb is to have a number of records n larger than 5(p+ 2), where p is the number of predictors.

• It can be shown that using predictors that are uncorrelated with the out- come variable increases the variance of predictions.

• It can be shown that dropping predictors that are actually correlated with the outcome variable can increase the average error (bias) of predictions.

The last two points mean that there is a trade-off between too few and too many predictors. In general, accepting some bias can reduce the variance in pre- dictions. This bias–variance trade-off is particularly important for large numbers of predictors, because in that case, it is very likely that there are variables in the model that have small coefficients relative to the standard deviation of the noise and also exhibit at least moderate correlation with other variables. Dropping such variables will improve the predictions, as it reduces the prediction variance. This type of bias–variance trade-off is a basic aspect of most data mining proce- dures for prediction and classification. In light of this, methods for reducing the number of predictors p to a smaller set are often used.

How to Reduce the Number of Predictors

The first step in trying to reduce the number of predictors should always be to use domain knowledge. It is important to understand what the various predictors are measuring and why they are relevant for predicting the outcome variable. With this knowledge, the set of predictors should be reduced to a sensible set that reflects the problem at hand. Some practical reasons for predictor elimination

VARIABLE SELECTION IN LINEAR REGRESSION 163

are the expense of collecting this information in the future; inaccuracy; high correlation with another predictor; many missing values; or simply irrelevance. Also helpful in examining potential predictors are summary statistics and graphs, such as frequency and correlation tables, predictor-specific summary statistics and plots, and missing value counts.

The next step makes use of computational power and statistical performance metrics. In general, there are two types of methods for reducing the number of predictors in a model. The first is an exhaustive search for the “best” subset of predictors by fitting regression models with all the possible combinations of predictors. The exhaustive search approach is not practical in many applications, and implementation in R can be tedious and unstable. The second approach is to search through a partial set of models. We describe these two approaches next.

Exhaustive Search The idea here is to evaluate all subsets of predictors. Since the number of subsets for even moderate values of p is very large, after the algorithm creates the subsets and runs all the models, we need some way to examine the most promising subsets and to select from them. The challenge is to select a model that is not too simplistic in terms of excluding important parameters (the model is under-fit), nor overly complex thereby modeling random noise (the model is over-fit). Several criteria for evaluating and comparing models are based on metrics computed from the training data:

One popular criterion is the adjusted R2, which is defined as

R2adj = 1− n− 1

n− p− 1 (1−R2),

where R2 is the proportion of explained variability in the model (in a model with a single predictor, this is the squared correlation). Like R2, higher values of R2adj indicate better fit. Unlike R

2, which does not account for the number of predictors used, R2adj uses a penalty on the number of predictors. This avoids the artificial increase in R2 that can result from simply increasing the number of predictors but not the amount of information. It can be shown that using R2adj to choose a subset is equivalent to picking the subset that minimizes σ̂2.

A second popular set of criteria for balancing under-fitting and over-fitting are the Akaike Information Criterion (AIC) and Schwartz’s Bayesian Information Criterion (BIC). AIC and BIC measure the goodness of fit of a model, but also include a penalty that is a function of the number of parameters in the model. As such, they can be used to compare various models for the same data set. AIC and BIC are estimates of prediction error based in information theory. Their derivation is beyond the scope of this book, but suffice it to say that models with smaller AIC and BIC values are considered better.

164 MULTIPLE LINEAR REGRESSION

A third criterion often used for subset selection is Mallow’s Cp (see formula below1). This criterion assumes that the full model (with all predictors) is unbi- ased, although it may have predictors that if dropped would reduce prediction variability. With this assumption, we can show that if a subset model is unbiased, the average Cp value equals p+ 1 (= number of predictors + 1), the size of the subset. So a reasonable approach to identifying subset models with small bias is to examine those with values of Cp that are near p+ 1. Good models are those that have values of Cp near p + 1 and that have small p (i.e., are of small size). Cp is computed from the formula

Cp = SSE

σ̂2full + 2(p+ 1)− n, (6.3)

where σ̂2full is the estimated value of σ 2 in the full model that includes all predic-

tors. It is important to remember that the usefulness of this approach depends heavily on the reliability of the estimate of σ2 for the full model. This requires that the training set contain a large number of records relative to the number of predictors. Note: It can be shown that for linear regression, in large samples Mallows’s Cp is equivalent to AIC.

Finally, a useful point to note is that for a fixed size of subset, R2, R2adj, Cp, AIC, and BIC all select the same subset. In fact, there is no difference between them in the order of merit they ascribe to subsets of a fixed size. This is good to know if comparing models with the same number of predictors, but often we want to compare models with different numbers of predictors.

Table 6.5 gives the results of applying an exhaustive search on the Toyota Corolla price data (with the 11 predictors). It reports the best model with a single predictor, two predictors, and so on. It can be seen that the R2adj increases until eight predictors are used (number of coefficients = 9) and then stabilizes. The Cp indicates that a model with 7 to 8 predictors is good. The dominant predictor in all models is the age of the car, with horsepower and mileage playing important roles as well.

Popular Subset Selection Algorithms The second method of finding the best subset of predictors relies on a partial, iterative search through the space of all possible regression models. The end product is one best subset of pre- dictors (although there do exist variations of these methods that identify several

1Mallow’s Cp is unrelated to the CP, or complexity parameter, used in classification and regression trees, described in Chapter 9).

VARIABLE SELECTION IN LINEAR REGRESSION 165

TABLE 6.5 EXHAUSTIVE SEARCH FOR REDUCING PREDICTORS IN TOYOTA COROLLA EXAMPLE

code for best subset

# use regsubsets() in package leaps to run an exhaustive search. # unlike with lm, categorical predictors must be turned into dummies manually. library(leaps) # create dummies for fuel type Fuel_Type <- as.data.frame(model.matrix(~ 0 + Fuel_Type, data=train.df)) # replace Fuel_Type column with 2 dummies train.df <- cbind(train.df[,-4], Fuel_Type[,]) head(train.df) search <- regsubsets(Price ~ ., data = train.df, nbest = 1, nvmax = dim(train.df)[2],

method = "exhaustive") sum <- summary(search)

# show models sum$which

# show metrics sum$rsq sum$adjr2 sum$Cp

Output

> sum$which (Intercept) Age_08_04 KM HP Met_Color Automatic CC Doors Quarterly_Tax

1 TRUE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE 2 TRUE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE 3 TRUE TRUE FALSE TRUE FALSE FALSE FALSE FALSE FALSE 4 TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE FALSE 5 TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE TRUE 6 TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE TRUE 7 TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE TRUE 8 TRUE TRUE TRUE TRUE FALSE TRUE FALSE FALSE TRUE 9 TRUE TRUE TRUE TRUE FALSE TRUE FALSE TRUE TRUE 10 TRUE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE 11 TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE

Weight Fuel_TypeCNG Fuel_TypeDiesel Fuel_TypePetrol 1 FALSE FALSE FALSE FALSE 2 TRUE FALSE FALSE FALSE 3 TRUE FALSE FALSE FALSE 4 TRUE FALSE FALSE FALSE 5 TRUE FALSE FALSE FALSE 6 TRUE FALSE FALSE TRUE 7 TRUE TRUE TRUE FALSE 8 TRUE TRUE TRUE FALSE 9 TRUE TRUE TRUE FALSE 10 TRUE FALSE TRUE TRUE 11 TRUE TRUE TRUE FALSE

> sum$rsq [1] 0.7560439 0.7929293 0.8276610 0.8447333 0.8506850 0.8549587 [7] 0.8561788 0.8564857 0.8565820 0.8566933 0.8567187

> sum$adjr2 [1] 0.7556359 0.7922356 0.8267935 0.8436895 0.8494282 0.8534911 [7] 0.8544782 0.8545430 0.8543943 0.8542602 0.8540382

> sum$cp [1] 403.45 254.33 114.04 46.10 23.72 8.21 5.21 5.95 7.56 9.10 11.00

166 MULTIPLE LINEAR REGRESSION

close-to-best choices for different sizes of predictor subsets). This approach is computationally cheaper, but it has the potential of missing “good” combina- tions of predictors. None of the methods guarantee that they yield the best subset for any criterion, such as R2adj. They are reasonable methods for situations with a large number of predictors, but for a moderate number of predictors, the exhaustive search is preferable.

Three popular iterative search algorithms are forward selection, backward elimi- nation, and stepwise regression. In forward selection, we start with no predictors and then add predictors one by one. Each predictor added is the one (among all predictors) that has the largest contribution to R2 on top of the predictors that are already in it. The algorithm stops when the contribution of additional pre- dictors is not statistically significant. The main disadvantage of this method is that the algorithm will miss pairs or groups of predictors that perform very well together but perform poorly as single predictors. This is similar to interviewing job candidates for a team project one by one, thereby missing groups of candi- dates who perform superiorly together (“colleagues”), but poorly on their own or with non-colleagues.

In backward elimination, we start with all predictors and then at each step, eliminate the least useful predictor (according to statistical significance). The algorithm stops when all the remaining predictors have significant contributions. The weakness of this algorithm is that computing the initial model with all predictors can be time-consuming and unstable. Stepwise regression is like forward selection except that at each step, we consider dropping predictors that are not statistically significant, as in backward elimination.

R has several libraries with stepwise functions: function regsubsets() in the leaps package implements (in addition to exhaustive search) forward selection, backward elimination, and stepwise regression. Predictors are added/dropped based on eitherR2,R2adj, orCp. In contrast, function step() in the stats package, as well as function stepAIC() in the MASS package perform model selection using the AIC criterion (stepAIC offers a wider range of object classes).

Table 6.6 shows the result of backward elimination for the Toyota Corolla example. The chosen seven-predictor model is identical to the best seven- predictor model chosen by the exhaustive search. However, recall that the exhaustive search indicated a higher R2adj for the eight-predictor model. In com- parison, forward selection (Table 6.7) selected an 11-predictor model, thereby not eliminating any predictor. The results for stepwise selection, seen in Table 6.8, are the same as those obtained by backward elimination.

Finally, additional ways to reduce the dimension of the data are by using principal components (Chapter 4) and regression trees (Chapter 9).

VARIABLE SELECTION IN LINEAR REGRESSION 167

TABLE 6.6 BACKWARD ELIMINATION FOR REDUCING PREDICTORS IN TOYOTA COROLLA EXAMPLE

code for stepwise regression

# use step() to run stepwise regression. # set directions = to either "backward", "forward", or "both". car.lm.step <- step(car.lm, direction = "backward") summary(car.lm.step) # Which variables did it drop? car.lm.step.pred <- predict(car.lm.step, valid.df) accuracy(car.lm.step.pred, valid.df$Price)

Output

> summary(car.lm.step)

Call: lm(formula = Price ~ Age_08_04 + KM + Fuel_Type + HP + Quarterly_Tax +

Weight, data = train.df)

Residuals: Min 1Q Median 3Q Max

-8263 -825 1 839 7312

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) -1853.36897 1620.35672 -1.14 0.253 Age_08_04 -135.72630 4.83995 -28.04 < 0.0000000000000002 *** KM -0.01912 0.00233 -8.19 0.0000000000000016 *** Fuel_TypeDiesel 1179.35368 526.25097 2.24 0.025 * Fuel_TypePetrol 2374.05722 517.80593 4.58 0.0000055461532557 *** HP 39.27366 5.51783 7.12 0.0000000000031903 *** Quarterly_Tax 16.43837 2.58633 6.36 0.0000000004140248 *** Weight 12.74441 1.47320 8.65 < 0.0000000000000002 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1400 on 592 degrees of freedom Multiple R-squared: 0.856, Adjusted R-squared: 0.854 F-statistic: 503 on 7 and 592 DF, p-value: <0.0000000000000002

> car.lm.step.pred <- predict(car.lm.step, valid.df) > accuracy(car.lm.step.pred, valid.df$Price)

ME RMSE MAE MPE MAPE Test set -38.9 1321 1016 -1.67 9.05

168 MULTIPLE LINEAR REGRESSION

TABLE 6.7 FORWARD SELECTION FOR REDUCING PREDICTORS IN TOYOTA COROLLA EXAMPLE

> summary(car.lm.step)

Call: lm(formula = Price ~ Age_08_04 + KM + Fuel_Type + HP + Met_Color +

Automatic + CC + Doors + Quarterly_Tax + Weight, data = train.df)

Residuals: Min 1Q Median 3Q Max

-8213 -839 -14 831 7271

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) -1774.87783 1643.74482 -1.08 0.281 Age_08_04 -135.43088 4.87591 -27.78 < 0.0000000000000002 *** KM -0.01900 0.00234 -8.12 0.0000000000000028 *** Fuel_TypeDiesel 1208.33916 534.43140 2.26 0.024 * Fuel_TypePetrol 2425.87671 520.58798 4.66 0.0000039169767967 *** HP 38.98554 5.58718 6.98 0.0000000000081162 *** Met_Color 84.79272 126.88345 0.67 0.504 Automatic 306.68415 289.43314 1.06 0.290 CC 0.03197 0.09908 0.32 0.747 Doors -44.15774 64.05653 -0.69 0.491 Quarterly_Tax 16.67734 2.60267 6.41 0.0000000003028702 *** Weight 12.66749 1.53659 8.24 0.0000000000000011 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1410 on 588 degrees of freedom Multiple R-squared: 0.857, Adjusted R-squared: 0.854 F-statistic: 320 on 11 and 588 DF, p-value: <0.0000000000000002

TABLE 6.8 STEPWISE REGRESSION FOR REDUCING PREDICTORS IN TOYOTA COROLLA EXAMPLE

> summary(car.lm.step)

Call: lm(formula = Price ~ Age_08_04 + KM + Fuel_Type + HP + Quarterly_Tax +

Weight, data = train.df)

Residuals: Min 1Q Median 3Q Max

-8263 -825 1 839 7312

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) -1853.36897 1620.35672 -1.14 0.253 Age_08_04 -135.72630 4.83995 -28.04 < 0.0000000000000002 *** KM -0.01912 0.00233 -8.19 0.0000000000000016 *** Fuel_TypeDiesel 1179.35368 526.25097 2.24 0.025 * Fuel_TypePetrol 2374.05722 517.80593 4.58 0.0000055461532557 *** HP 39.27366 5.51783 7.12 0.0000000000031903 *** Quarterly_Tax 16.43837 2.58633 6.36 0.0000000004140248 *** Weight 12.74441 1.47320 8.65 < 0.0000000000000002 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 1400 on 592 degrees of freedom Multiple R-squared: 0.856, Adjusted R-squared: 0.854 F-statistic: 503 on 7 and 592 DF, p-value: <0.0000000000000002

PROBLEMS 169

PROBLEMS

6.1 Predicting Boston Housing Prices. The file BostonHousing.csv contains informa- tion collected by the US Bureau of the Census concerning housing in the area of Boston, Massachusetts. The dataset includes information on 506 census housing tracts in the Boston area. The goal is to predict the median house price in new tracts based on information such as crime rate, pollution, and number of rooms. The dataset con- tains 13 predictors, and the response is the median house price (MEDV). Table 6.9 describes each of the predictors and the response.

TABLE 6.9 DESCRIPTION OF VARIABLES FOR BOSTON HOUSING EXAMPLE

CRIM Per capita crime rate by town ZN Proportion of residential land zoned for lots over 25,000 ft2

INDUS Proportion of nonretail business acres per town CHAS Charles River dummy variable (= 1 if tract bounds river; = 0 otherwise) NOX Nitric oxide concentration (parts per 10 million) RM Average number of rooms per dwelling AGE Proportion of owner-occupied units built prior to 1940 DIS Weighted distances to five Boston employment centers RAD Index of accessibility to radial highways TAX Full-value property-tax rate per $10,000 PTRATIO Pupil/teacher ratio by town LSTAT Percentage lower status of the population MEDV Median value of owner-occupied homes in $1000s

a. Why should the data be partitioned into training and validation sets? What will the training set be used for? What will the validation set be used for?

b. Fit a multiple linear regression model to the median house price (MEDV) as a function of CRIM, CHAS, and RM. Write the equation for predicting the median house price from the predictors in the model.

c. Using the estimated regression model, what median house price is predicted for a tract in the Boston area that does not bound the Charles River, has a crime rate of 0.1, and where the average number of rooms per house is 6? What is the prediction error?

d. Reduce the number of predictors:

i. Which predictors are likely to be measuring the same thing among the 13 predictors? Discuss the relationships among INDUS, NOX, and TAX.

ii. Compute the correlation table for the 12 numerical predictors and search for highly correlated pairs. These have potential redundancy and can cause multi- collinearity. Choose which ones to remove based on this table.

iii. Use stepwise regression with the three options (backward, forward, both) to reduce the remaining predictors as follows: Run stepwise on the training set. Choose the top model from each stepwise run. Then use each of these models separately to predict the validation set. Compare RMSE, MAPE, and mean error, as well as lift charts. Finally, describe the best model.

6.2 Predicting Software Reselling Profits. Tayko Software is a software catalog firm that sells games and educational software. It started out as a software manufacturer

170 MULTIPLE LINEAR REGRESSION

and then added third-party titles to its offerings. It recently revised its collection of items in a new catalog, which it mailed out to its customers. This mailing yielded 2000 purchases. Based on these data, Tayko wants to devise a model for predicting the spending amount that a purchasing customer will yield. The file Tayko.csv contains information on 2000 purchases. Table 6.10 describes the variables to be used in the problem (the Excel file contains additional variables).

TABLE 6.10 DESCRIPTION OF VARIABLES FOR TAYKO SOFTWARE EXAMPLE

FREQ Number of transactions in the preceding year LAST_UPDATE Number of days since last update to customer record WEB Whether customer purchased by Web order at least once GENDER Male or female ADDRESS_RES Whether it is a residential address ADDRESS_US Whether it is a US address SPENDING (response) Amount spent by customer in test mailing (in dollars)

a. Explore the spending amount by creating a pivot table for the categorical variables and computing the average and standard deviation of spending in each category.

b. Explore the relationship between spending and each of the two continuous predictors by creating two scatterplots (Spending vs. Freq, and Spending vs. last_update_days_ago. Does there seem to be a linear relationship?

c. To fit a predictive model for Spending:

i. Partition the 2000 records into training and validation sets.

ii. Run a multiple linear regression model for Spending vs. all six predictors. Give the estimated predictive equation.

iii. Based on this model, what type of purchaser is most likely to spend a large amount of money?

iv. If we used backward elimination to reduce the number of predictors, which predictor would be dropped first from the model?

v. Show how the prediction and the prediction error are computed for the first purchase in the validation set.

vi. Evaluate the predictive accuracy of the model by examining its performance on the validation set.

vii. Create a histogram of the model residuals. Do they appear to follow a normal distribution? How does this affect the predictive performance of the model?

6.3 Predicting Airfare on New Routes. The following problem takes place in the United States in the late 1990s, when many major US cities were facing issues with airport congestion, partly as a result of the 1978 deregulation of airlines. Both fares and routes were freed from regulation, and low-fare carriers such as Southwest (SW) began competing on existing routes and starting non- stop service on routes that previously lacked it. Building completely new airports is generally not feasible, but sometimes decommissioned military bases or smaller munic- ipal airports can be reconfigured as regional or larger commercial airports. There are numerous players and interests involved in the issue (airlines, city, state and federal authorities, civic groups, the military, airport operators), and an aviation consulting firm is seeking advisory contracts with these players. The firm needs predictive models

PROBLEMS 171

to support its consulting service. One thing the firm might want to be able to predict is fares, in the event a new airport is brought into service. The firm starts with the file Airfares.csv, which contains real data that were collected between Q3-1996 and Q2-1997. The variables in these data are listed in Table 6.11, and are believed to be important in predicting FARE. Some airport-to-airport data are available, but most data are at the city-to-city level. One question that will be of interest in the analysis is the effect that the presence or absence of Southwest has on FARE.

TABLE 6.11 DESCRIPTION OF VARIABLES FOR AIRFARE EXAMPLE

S_CODE Starting airport’s code S_CITY Starting city E_CODE Ending airport’s code E_CITY Ending city COUPON Average number of coupons (a one-coupon flight is a nonstop flight,

a two-coupon flight is a one-stop flight, etc.) for that route NEW Number of new carriers entering that route between Q3-96 and Q2-97 VACATION Whether (Yes) or not (No) a vacation route SW Whether (Yes) or not (No) Southwest Airlines serves that route HI Herfindahl index: measure of market concentration S_INCOME Starting city’s average personal income E_INCOME Ending city’s average personal income S_POP Starting city’s population E_POP Ending city’s population SLOT Whether or not either endpoint airport is slot-controlled

(this is a measure of airport congestion) GATE Whether or not either endpoint airport has gate constraints

(this is another measure of airport congestion) DISTANCE Distance between two endpoint airports in miles PAX Number of passengers on that route during period of data collection FARE Average fare on that route

a. Explore the numerical predictors and response (FARE) by creating a correlation table and examining some scatterplots between FARE and those predictors. What seems to be the best single predictor of FARE?

b. Explore the categorical predictors (excluding the first four) by computing the per- centage of flights in each category. Create a pivot table with the average fare in each category. Which categorical predictor seems best for predicting FARE?

c. Find a model for predicting the average fare on a new route:

i. Convert categorical variables (e.g., SW) into dummy variables. Then, partition the data into training and validation sets. The model will be fit to the training data and evaluated on the validation set.

ii. Use stepwise regression to reduce the number of predictors. You can ignore the first four predictors (S_CODE, S_CITY, E_CODE, E_CITY). Report the estimated model selected.

iii. Repeat (ii) using exhaustive search instead of stepwise regression. Compare the resulting best model to the one you obtained in (ii) in terms of the predictors that are in the model.

iv. Compare the predictive accuracy of both models (ii) and (iii) using measures such as RMSE and average error and lift charts.

172 MULTIPLE LINEAR REGRESSION

v. Using model (iii), predict the average fare on a route with the following char- acteristics: COUPON = 1.202, NEW = 3, VACATION = No, SW = No, HI = 4442.141, S_INCOME = $28,760, E_INCOME = $27,664, S_POP = 4,557,004, E_POP = 3,195,503, SLOT = Free, GATE = Free, PAX = 12,782, DISTANCE = 1976 miles.

vi. Predict the reduction in average fare on the route in (v) if Southwest decides to cover this route [using model (iii)].

vii. In reality, which of the factors will not be available for predicting the average fare from a new airport (i.e., before flights start operating on those routes)? Which ones can be estimated? How?

viii. Select a model that includes only factors that are available before flights begin to operate on the new route. Use an exhaustive search to find such a model.

ix. Use the model in (viii) to predict the average fare on a route with character- istics COUPON = 1.202, NEW = 3, VACATION = No, SW = No, HI = 4442.141, S_INCOME = $28,760, E_INCOME = $27,664, S_ POP = 4,557,004, E_POP = 3,195,503, SLOT = Free, GATE = Free, PAX = 12782, DISTANCE = 1976 miles.

x. Compare the predictive accuracy of this model with model (iii). Is this model good enough, or is it worthwhile reevaluating the model once flights begin on the new route?

d. In competitive industries, a new entrant with a novel business plan can have a disruptive effect on existing firms. If a new entrant’s business model is sustainable, other players are forced to respond by changing their business practices. If the goal of the analysis was to evaluate the effect of Southwest Airlines’ presence on the airline industry rather than predicting fares on new routes, how would the analysis be different? Describe technical and conceptual aspects.

6.4 Predicting Prices of Used Cars. The file ToyotaCorolla.csv contains data on used cars (Toyota Corolla) on sale during late summer of 2004 in the Netherlands. It has 1436 records containing details on 38 attributes, including Price, Age, Kilometers, HP, and other specifications. The goal is to predict the price of a used Toyota Corolla based on its specifications. (The example in Section 6.3 is a subset of this dataset.)

Split the data into training (50%), validation (30%), and test (20%) datasets. Run a multiple linear regression with the outcome variable Price and pre-

dictor variables Age_08_04, KM, Fuel_Type, HP, Automatic, Doors, Quar- terly_Tax, Mfr_Guarantee, Guarantee_Period, Airco, Automatic_airco, CD_Player, Powered_Windows, Sport_Model, and Tow_Bar.

a. What appear to be the three or four most important car specifications for predicting the car’s price?

b. Using metrics you consider useful, assess the performance of the model in predict- ing prices.

CHAPTER 7

k-Nearest Neighbors (k-NN)

In this chapter, we describe the k-nearest-neighbors algorithm that can be used for classification (of a categorical outcome) or prediction (of a numerical out- come). To classify or predict a new record, the method relies on finding “similar” records in the training data. These “neighbors” are then used to derive a classifi- cation or prediction for the new record by voting (for classification) or averaging (for prediction). We explain how similarity is determined, how the number of neighbors is chosen, and how a classification or prediction is computed. k-NN is a highly automated data-driven method. We discuss the advantages and weak- nesses of the k-NN method in terms of performance and practical considerations such as computational time.

7.1 The k-NN Classifier (Categorical Outcome)

The idea in k-nearest-neighbors methods is to identify k records in the training dataset that are similar to a new record that we wish to classify. We then use these similar (neighboring) records to classify the new record into a class, assigning the new record to the predominant class among these neighbors. Denote the values of the predictors for this new record by x1, x2, . . . , xp. We look for records in our training data that are similar or “near” the record to be classified in the predictor space (i.e., records that have values close to x1, x2, . . . , xp). Then, based on the classes to which those proximate records belong, we assign a class to the record that we want to classify.

Determining Neighbors

The k-nearest-neighbors algorithm is a classification method that does not make assumptions about the form of the relationship between the class membership

173

174 k-NEAREST NEIGHBORS (k-NN)

(Y ) and the predictors X1, X2, . . . , Xp. This is a nonparametric method because it does not involve estimation of parameters in an assumed function form, such as the linear form assumed in linear regression (Chapter 6). Instead, this method draws information from similarities between the predictor values of the records in the dataset.

A central question is how to measure the distance between records based on their predictor values. The most popular measure of distance is the Euclidean distance. The Euclidean distance between two records (x1, x2, . . . , xp) and (u1, u2, . . . , up) is√

(x1 − u1)2 + (x2 − u2)2 + · · ·+ (xp − up)2. (7.1)

You will find a host of other distance metrics in Chapters 12 and 15 for both numerical and categorical variables. However, the k-NN algorithm relies on many distance computations (between each record to be predicted and every record in the training set), and therefore the Euclidean distance, which is com- putationally cheap, is the most popular in k-NN.

To equalize the scales that the various predictors may have, note that in most cases, predictors should first be standardized before computing a Euclidean distance. Also note that the means and standard deviations used to standardize new records are those of the training data, and the new record is not included in calculating them. The validation data, like new data, are also not included in this calculation.

Classification Rule

After computing the distances between the record to be classified and existing records, we need a rule to assign a class to the record to be classified, based on the classes of its neighbors. The simplest case is k = 1, where we look for the record that is closest (the nearest neighbor) and classify the new record as belonging to the same class as its closest neighbor. It is a remarkable fact that this simple, intuitive idea of using a single nearest neighbor to classify records can be very powerful when we have a large number of records in our training set. It turns out that the misclassification error of the 1-nearest neighbor scheme has a misclassification rate that is no more than twice the error when we know exactly the probability density functions for each class.

The idea of the 1-nearest neighbor can be extended to k > 1 neighbors as follows:

1. Find the nearest k neighbors to the record to be classified.

2. Use a majority decision rule to classify the record, where the record is classified as a member of the majority class of the k neighbors.

THE k-NN CLASSIFIER (CATEGORICAL OUTCOME) 175

Example: Riding Mowers

A riding-mower manufacturer would like to find a way of classifying families in a city into those likely to purchase a riding mower and those not likely to buy one. A pilot random sample is undertaken of 12 owners and 12 nonowners in the city. The data are shown in Table 7.1. We first partition the data into training data (14 households) and validation data (10 households). Obviously, this dataset is too small for partitioning, which can result in unstable results, but we will continue with this partitioning for illustration purposes. A scatter plot of the training data is shown in Figure 7.1.

TABLE 7.1 LOT SIZE, INCOME, AND OWNERSHIP OF A RIDING MOWER FOR 24 HOUSEHOLDS

Household Income Lot Size Ownership of Number ($000s) (000s ft2) Riding Mower

1 60.0 18.4 Owner 2 85.5 16.8 Owner 3 64.8 21.6 Owner 4 61.5 20.8 Owner 5 87.0 23.6 Owner 6 110.1 19.2 Owner 7 108.0 17.6 Owner 8 82.8 22.4 Owner 9 69.0 20.0 Owner

10 93.0 20.8 Owner 11 51.0 22.0 Owner 12 81.0 20.0 Owner 13 75.0 19.6 Nonowner 14 52.8 20.8 Nonowner 15 64.8 17.2 Nonowner 16 43.2 20.4 Nonowner 17 84.0 17.6 Nonowner 18 49.2 17.6 Nonowner 19 59.4 16.0 Nonowner 20 66.0 18.4 Nonowner 21 47.4 16.4 Nonowner 22 33.0 18.8 Nonowner 23 51.0 14.0 Nonowner 24 63.0 14.8 Nonowner 25 60.0 20.0 ?

Now consider a new household with $60,000 income and lot size 20,000 ft2

(also shown in Figure 7.1). Among the households in the training set, the one closest to the new household (in Euclidean distance after normalizing income and lot size) is household 9, with $69,000 income and lot size 20,000 ft2. If we use a 1-NN classifier, we would classify the new household as an owner, like household 9. If we use k = 3, the three nearest households are 9, 14, and 1, as can be seen visually in the scatter plot, and as computed by the software (see output in Table 7.2). Two of these neighbors are owners of riding mowers,

176 k-NEAREST NEIGHBORS (k-NN)

code for loading and partitioning the riding mower data, and plotting scatter plot

mower.df <- read.csv("RidingMowers.csv") set.seed(111) train.index <- sample(row.names(mower.df), 0.6*dim(mower.df)[1]) valid.index <- setdiff(row.names(mower.df), train.index) train.df <- mower.df[train.index, ] valid.df <- mower.df[valid.index, ] ## new household new.df <- data.frame(Income = 60, Lot_Size = 20)

## scatter plot plot(Lot_Size ~ Income, data=train.df, pch=ifelse(train.df$Ownership=="Owner", 1, 3)) text(train.df$Income, train.df$Lot_Size, rownames(train.df), pos=4) text(60, 20, "X") legend("topright", c("owner", "non-owner", "newhousehold"), pch = c(1, 3, 4))

50 60 70 80 90 100 110

16 17

18 19

20 21

22

Income

Lo t_

S iz

e

15 17

9

11 8

201

10

7

2

19

14

18

12X

owner

nonowner

newhousehold

FIGURE 7.1 SCATTER PLOT OF LOT SIZE VS. INCOME FOR THE 18 HOUSEHOLDS IN THE TRAINING SET AND THE NEW HOUSEHOLD TO BE CLASSIFIED

and the last is a nonowner. The majority vote is therefore owner, and the new household would be classified as an owner (see bottom of output in Table 7.2).

Choosing k

The advantage of choosing k > 1 is that higher values of k provide smoothing that reduces the risk of overfitting due to noise in the training data. Generally speaking, if k is too low, we may be fitting to the noise in the data. However, if k is too high, we will miss out on the method’s ability to capture the local structure in the data, one of its main advantages. In the extreme, k = n = the number of records in the training dataset. In that case, we simply assign

THE k-NN CLASSIFIER (CATEGORICAL OUTCOME) 177

TABLE 7.2 RUNNING k-NN

code for normalizing data and finding nearest neighbors

# initialize normalized training, validation data, complete data frames to originals train.norm.df <- train.df valid.norm.df <- valid.df mower.norm.df <- mower.df # use preProcess() from the caret package to normalize Income and Lot_Size. norm.values <- preProcess(train.df[, 1:2], method=c("center", "scale")) train.norm.df[, 1:2] <- predict(norm.values, train.df[, 1:2]) valid.norm.df[, 1:2] <- predict(norm.values, valid.df[, 1:2]) mower.norm.df[, 1:2] <- predict(norm.values, mower.df[, 1:2]) new.norm.df <- predict(norm.values, new.df)

# use knn() to compute knn. # knn() is available in library FNN (provides a list of the nearest neighbors) # and library class (allows a numerical output variable). library(FNN) nn <- knn(train = train.norm.df[, 1:2], test = new.norm.df,

cl = train.norm.df[, 3], k = 3)

row.names(train.df)[attr(nn, "nn.index")]

Output

> row.names(train.df)[attr(nn, "nn.index")] [1] "9" "14" "1" < nn [1] Owner attr(,"nn.index") [,1][,2][,3] [1,] 3 12 7 attr(,"nn.dist")

[,1] [,2] [,3] [,1] 0.5137338 0.5716287 0.7946045 Levels: Owner

all records to the majority class in the training data, irrespective of the values of (x1, x2, . . . , xp), which coincides with the naive rule! This is clearly a case of oversmoothing in the absence of useful information in the predictors about the class membership. In other words, we want to balance between overfitting to the predictor information and ignoring this information completely. A bal- anced choice greatly depends on the nature of the data. The more complex and irregular the structure of the data, the lower the optimum value of k. Typically, values of k fall in the range 1 to 20. We will use odd numbers to avoid ties.

So how is k chosen? Answer: We choose the k with the best classification performance. We use the training data to classify the records in the validation data, then compute error rates for various choices of k. For our example, if we choose k = 1, we will classify in a way that is very sensitive to the local characteristics of the training data. On the other hand, if we choose a large

178 k-NEAREST NEIGHBORS (k-NN)

TABLE 7.3 ACCURACY (OR CORRECT RATE) OF k-NN PREDICTIONS IN VALIDATION SET FOR VARIOUS CHOICES OF k.

code for measuring the accuracy of different k values

library(caret)

# initialize a data frame with two columns: k, and accuracy. accuracy.df <- data.frame(k = seq(1, 14, 1), accuracy = rep(0, 14))

# compute knn for different k on validation. for(i in 1:14) {

knn.pred <- knn(train.norm.df[, 1:2], valid.norm.df[, 1:2], cl = train.norm.df[, 3], k = i)

accuracy.df[i, 2] <- confusionMatrix(knn.pred, valid.norm.df[, 3])$overall[1] }

Output

> accuracy.df k accuracy

1 1 0.7 2 2 0.7 3 3 0.8 4 4 0.9 5 5 0.8 6 6 0.9 7 7 0.9 8 8 1.0 9 9 0.9 10 10 0.9 11 11 0.9 12 12 0.8 13 13 0.4 14 14 0.4

value of k, such as k = 18, we would simply predict the most frequent class in the dataset in all cases. This is a very stable prediction but it completely ignores the information in the predictors. To find a balance, we examine the accuracy (of predictions in the validation set) that results from different choices of k between 1 and 14. For an even number k, if there is a tie in classifying a household, the tie is broken randomly.1 This is shown in Table 7.3. We would choose k = 4, which maximizes our accuracy in the validation set.2 Note, however, that now the validation set is used as part of the training process (to set k) and does not reflect a true holdout set as before. Ideally, we would want a third test set to evaluate the performance of the method on data that it did not see.

1If you are interested in reproducibility of results, check the accuracy only for each odd k. 2Partitioning such a small dataset is unwise in practice, as results will heavily rely on the particular partition. For instance, if you use a different partitioning, you might obtain a different “optimal” k. We use this example for illustration only.

THE k-NN CLASSIFIER (CATEGORICAL OUTCOME) 179

TABLE 7.4 CLASSIFYING A NEW HOUSEHOLD USING THE “BEST K” = 4

code for running the k-NN algorithm to classify the new household

knn.pred.new <- knn(mower.norm.df[, 1:2], new.norm.df, cl = mower.norm.df[, 3], k = 4)

row.names(train.df)[attr(nn, "nn.index")]

Partial Output

> row.names(train.df)[attr(nn, "nn.index")] [1] "9" "14" "1" "20" > knn.pred.new [1] Owner attr(,"nn.index")

[,1] [,2] [,3] [,4] [1,] 3 12 7 6 attr(,"nn.dist")

[,1] [,2] [,3] [,4] [1,] 0.514 0.572 0.795 0.865 Levels: Owner

Once k is chosen, we rerun the algorithm on the combined training and testing sets in order to generate classifications of new records. An example is shown in Table 7.4, where the four nearest neighbors are used to classify the new household.

Setting the Cutoff Value

k-NN uses a majority decision rule to classify a new record, where the record is classified as a member of the majority class of the k neighbors. The definition of “majority” is directly linked to the notion of a cutoff value applied to the class membership probabilities. Let us consider a binary outcome case. For a new record, the proportion of class 1 members among its neighbors is an estimate of its propensity (probability) of belonging to class 1. In the riding mowers example with k = 4, we found that the four nearest neighbors to the new household (with income = $60,000 and lot size = 20,000 ft2) are households 9, 14, 1, and 20. Since two of these are owners and the other two are nonowners, we can estimate for the new household a probability of 0.5 of being an owner (and 0.5 for being a nonowner). Using a simple majority rule is equivalent to setting the cutoff value to 0.5. In Table 7.4, we see that the software broke the tie by (randomly) assigning class owner to this record.

As mentioned in Chapter 5, changing the cutoff value affects the confusion matrix (i.e., the error rates). Hence, in some cases we might want to choose a cutoff other than the default 0.5 for the purpose of maximizing accuracy or for incorporating misclassification costs.

180 k-NEAREST NEIGHBORS (k-NN)

k-NN with More Than Two Classes

The k-NN classifier can easily be applied to an outcome with m classes, where m > 2. The “majority rule” means that a new record is classified as a member of the majority class of its k neighbors. An alternative, when there is a specific class that we are interested in identifying (and are willing to “overidentify” records as belonging to this class), is to calculate the proportion of the k neighbors that belong to this class of interest, use that as an estimate of the probability (propensity) that the new record belongs to that class, and then refer to a user- specified cutoff value to decide whether to assign the new record to that class. For more on the use of cutoff value in classification where there is a single class of interest, see Chapter 5.

Converting Categorical Variables to Binary Dummies

It usually does not make sense to calculate Euclidean distance between two non- numeric categories (e.g., cookbooks and maps, in a bookstore). Therefore, before k-NN can be applied, categorical variables must be converted to binary dummies. In contrast to the situation with statistical models such as regression, allm binaries should be created and used with k-NN. While mathematically this is redundant, sincem−1 dummies contain the same information asm dummies, this redundant information does not create the multicollinearity problems that it does for linear models. Moreover, in k-NN the use of m−1 dummies can yield different classifications than the use of m dummies, and lead to an imbalance in the contribution of the different categories to the model.

7.2 k-NN for a Numerical Outcome

The idea of k-NN can readily be extended to predicting a continuous value (as is our aim with multiple linear regression models). The first step of determining neighbors by computing distances remains unchanged. The second step, where a majority vote of the neighbors is used to determine class, is modified such that we take the average outcome value of the k-nearest neighbors to determine the prediction. Often, this average is a weighted average, with the weight decreasing with increasing distance from the point at which the prediction is required. In R, we can use function knn() in the class package to compute k-NN numerical predictions for the validation set.

Another modification is in the error metric used for determining the “best k.” Rather than the overall error rate used in classification, RMS error or another prediction error metric should be used in prediction (see Chapter 5).

k-NN FOR A NUMERICAL OUTCOME 181

P A N D O R A

Pandora is an Internet music radio service that allows users to build customized “stations” that play music similar to a song or artist that they have specified. Pan- dora uses a k-NN style clustering/classification process called the Music Genome Project to locate new songs or artists that are close to the user-specified song or artist.

Pandora was the brainchild of Tim Westergren, who worked as a musician and a nanny when he graduated from Stanford in the 1980s. Together with Nolan Gasser, who was studying medieval music, he developed a “matching engine” by entering data about a song’s characteristics into a spreadsheet. The first result was surprising— a Beatles song matched to a Bee Gees song, but they built a company around the concept. The early days were hard—Westergren racked up over $300,000 in personal debt, maxed out 11 credit cards, and ended up in the hospital once due to stress- induced heart palpitations. A venture capitalist finally invested funds in 2004 to rescue the firm, and as of 2013, it is listed on the NY Stock Exchange.

In simplified terms, the process works roughly as follows for songs:

1. Pandora has established hundreds of variables on which a song can be measured on a scale from 0–5. Four such variables from the beginning of the list are

• Acid Rock Qualities • Accordion Playing • Acousti-Lectric Sonority • Acousti-Synthetic Sonority

2. Pandora pays musicians to analyze tens of thousands of songs, and rate each song on each of these attributes. Each song will then be represented by a row vector of values between 0 and 5, for example, for Led Zeppelin’s Kashmir:

Kashmir 4 0 3 3 … (high on acid rock attributes, no accordion, etc.)

This step represents a costly investment, and lies at the heart of Pandora’s value because these variables have been tested and selected because they accurately reflect the essence of a song, and provide a basis for defining highly individu- alized preferences.

3. The online user specifies a song that s/he likes (the song must be in Pandora’s database).

4. Pandora then calculates the statistical distance1 between the user’s song, and the songs in its database. It selects a song that is close to the user-specified song and plays it.

5. The user then has the option of saying “I like this song,” “I don’t like this song,” or saying nothing.

6. If “like” is chosen, the original song, plus the new song are merged into a 2- song cluster2 that is represented by a single vector, comprised of means of the variables in the original two song vectors.

182 k-NEAREST NEIGHBORS (k-NN)

7. If “dislike” is chosen, the vector of the song that is not liked is stored for future reference. (If the user does not express an opinion about the song, in our simplified example here, the new song is not used for further comparisons.)

8. Pandora looks in its database for a new song, one whose statistical distance is close to the “like” song cluster,3 and not too close to the “dislike” song. Depending on the user’s reaction, this new song might be added to the “like” cluster or “dislike” cluster.

Over time, Pandora develops the ability to deliver songs that match a particular taste of a particular user. A single user might build up multiple stations around different song clusters. Clearly, this is a less limiting approach than selecting music in terms of which “genre” it belongs to.

While the process described above is a bit more complex than the basic “classi- fication of new data” process described in this chapter, the fundamental process— classifying a record according to its proximity to other records—is the same at its core. Note the role of domain knowledge in this machine learning process—the vari- ables have been tested and selected by the project leaders, and the measurements have been made by human experts.

Further reading: See www.pandora.com, Wikipedia’s article on the Music Genome Project, and Joyce John’s article “Pandora and the Music Genome Project,” Scientific Computing, vol. 23, no. 10: 14, p. 40–41, Sep. 2006.

1See Section 12.5 in Chapter 12 for an explanation of statistical distance. 2See Chapter 15 for more on clusters. 3See Case 21.6 “Segmenting Consumers of Bath Soap” for an exercise involving the identification of clusters, which are then used for classification purposes.

7.3 Advantages and Shortcomings of k-NN Algorithms

The main advantage of k-NN methods is their simplicity and lack of paramet- ric assumptions. In the presence of a large enough training set, these methods perform surprisingly well, especially when each class is characterized by multiple combinations of predictor values. For instance, in real-estate databases, there are likely to be multiple combinations of {home type, number of rooms, neighbor- hood, asking price, etc.} that characterize homes that sell quickly vs. those that remain for a long period on the market.

There are three difficulties with the practical exploitation of the power of the k-NN approach. First, although no time is required to estimate parameters from the training data (as would be the case for parametric models such as regression), the time to find the nearest neighbors in a large training set can be prohibitive. A

ADVANTAGES AND SHORTCOMINGS OF k-NN ALGORITHMS 183

number of ideas have been implemented to overcome this difficulty. The main ideas are:

• Reduce the time taken to compute distances by working in a reduced dimension using dimension reduction techniques such as principal com- ponents analysis (Chapter 4).

• Use sophisticated data structures such as search trees to speed up identifi- cation of the nearest neighbor. This approach often settles for an “almost nearest” neighbor to improve speed. An example is using bucketing, where the records are grouped into buckets so that records within each bucket are close to each other. For a to-be-predicted record, buckets are ordered by their distance to the record. Starting from the nearest bucket, the dis- tance to each of the records within the bucket is measured. The algorithm stops when the distance to a bucket is larger than the distance to the closest record thus far.

Second, the number of records required in the training set to qualify as large increases exponentially with the number of predictors p. This is because the expected distance to the nearest neighbor goes up dramatically with p unless the size of the training set increases exponentially with p. This phenomenon is known as the curse of dimensionality, a fundamental issue pertinent to all clas- sification, prediction, and clustering techniques. This is why we often seek to reduce the number of predictors through methods such as selecting subsets of the predictors for our model or by combining them using methods such as prin- cipal components analysis, singular value decomposition, and factor analysis (see Chapter 4).

Third, k-NN is a “lazy learner”: the time-consuming computation is deferred to the time of prediction. For every record to be predicted, we compute its distances from the entire set of training records only at the time of prediction. This behavior prohibits using this algorithm for real-time prediction of a large number of records simultaneously.

184 k-NEAREST NEIGHBORS (k-NN)

PROBLEMS

7.1 Calculating Distance with Categorical Predictors. This exercise with a tiny dataset illustrates the calculation of Euclidean distance, and the creation of binary dummies. The online education company Statistics.com segments its customers and prospects into three main categories: IT professionals (IT), statisticians (Stat), and other (Other). It also tracks, for each customer, the number of years since first contact (years). Consider the following customers; information about whether they have taken a course or not (the outcome to be predicted) is included:

Customer 1: Stat, 1 year, did not take course Customer 2: Other, 1.1 year, took course

a. Consider now the following new prospect: Prospect 1: IT, 1 year

Using the above information on the two customers and one prospect, create one dataset for all three with the categorical predictor variable transformed into 2 bina- ries, and a similar dataset with the categorical predictor variable transformed into 3 binaries.

b. For each derived dataset, calculate the Euclidean distance between the prospect and each of the other two customers. (Note: while it is typical to normalize data for k- NN, this is not an iron-clad rule and you may proceed here without normalization.)

c. Using k-NN with k = 1, classify the prospect as taking or not taking a course using each of the two derived datasets. Does it make a difference whether you use 2 or 3 dummies?

7.2 Personal Loan Acceptance. Universal Bank is a relatively young bank growing rapidly in terms of overall customer acquisition. The majority of these customers are liability customers (depositors) with varying sizes of relationship with the bank. The customer base of asset customers (borrowers) is quite small, and the bank is interested in expanding this base rapidly to bring in more loan business. In particular, it wants to explore ways of converting its liability customers to personal loan customers (while retaining them as depositors).

A campaign that the bank ran last year for liability customers showed a healthy conversion rate of over 9% success. This has encouraged the retail marketing depart- ment to devise smarter campaigns with better target marketing. The goal is to use k-NN to predict whether a new customer will accept a loan offer. This will serve as the basis for the design of a new campaign.

The file UniversalBank.csv contains data on 5000 customers. The data include customer demographic information (age, income, etc.), the customer’s relationship with the bank (mortgage, securities account, etc.), and the customer response to the last personal loan campaign (Personal Loan). Among these 5000 customers, only 480 (= 9.6%) accepted the personal loan that was offered to them in the earlier campaign.

Partition the data into training (60%) and validation (40%) sets.

a. Consider the following customer: Age = 40, Experience = 10, Income = 84, Family = 2, CCAvg = 2, Education_1 = 0, Education_2 = 1, Education_3 = 0, Mortgage = 0, Securities Account = 0, CD Account = 0, Online = 1, and Credit Card = 1. Perform a k-NN classification with all predictors except ID and ZIP code using k = 1. Remember to transform categorical predictors with more than two categories into dummy variables first.

PROBLEMS 185

Specify the success class as 1 (loan acceptance), and use the default cutoff value of 0.5. How would this customer be classified?

b. What is a choice of k that balances between overfitting and ignoring the predictor information?

c. Show the confusion matrix for the validation data that results from using the best k.

d. Consider the following customer: Age = 40, Experience = 10, Income = 84, Family = 2, CCAvg = 2, Education_1 = 0, Education_2 = 1, Education_3 = 0, Mortgage = 0, Securities Account = 0, CD Account = 0, Online = 1 and Credit Card = 1. Classify the customer using the best k.

e. Repartition the data, this time into training, validation, and test sets (50% : 30% : 20%). Apply the k-NN method with the k chosen above. Compare the confusion matrix of the test set with that of the training and validation sets. Comment on the differences and their reason.

7.3 Predicting Housing Median Prices. The file BostonHousing.csv contains infor- mation on over 500 census tracts in Boston, where for each tract multiple variables are recorded. The last column (CAT.MEDV) was derived from MEDV, such that it obtains the value 1 if MEDV > 30 and 0 otherwise. Consider the goal of predicting the median value (MEDV) of a tract, given the information in the first 12 columns.

Partition the data into training (60%) and validation (40%) sets.

a. Perform a k-NN prediction with all 12 predictors (ignore the CAT.MEDV col- umn), trying values of k from 1 to 5. Make sure to normalize the data, and choose function knn() from package class rather than package FNN. To make sure R is using the class package (when both packages are loaded), use class::knn(). What is the best k? What does it mean?

b. Predict the MEDV for a tract with the following information, using the best k:

CRIM ZN INDUS CHAS NOX RM AGE DIS RAD TAX PTRATIO LSTAT 0.2 0 7 0 0.538 6 62 4.7 4 307 21 10

c. If we used the above k-NN algorithm to score the training data, what would be the error of the training set?

d. Why is the validation data error overly optimistic compared to the error rate when applying this k-NN predictor to new data?

e. If the purpose is to predict MEDV for several thousands of new tracts, what would be the disadvantage of using k-NN prediction? List the operations that the algo- rithm goes through in order to produce each prediction.

CHAPTER 8

The Naive Bayes Classifier

In this chapter, we introduce the naive Bayes classifier, which can be applied to data with categorical predictors. We review the concept of conditional prob- abilities, then present the complete, or exact, Bayesian classifier. We next see how it is impractical in most cases, and learn how to modify it and use instead the naive Bayes classifier, which is more generally applicable.

8.1 Introduction

The naive Bayes method (and, indeed, an entire branch of statistics) is named after the Reverend Thomas Bayes (1702–1761). To understand the naive Bayes classifier, we first look at the complete, or exact, Bayesian classifier. The basic principle is simple. For each record to be classified:

1. Find all the other records with the same predictor profile (i.e., where the predictor values are the same).

2. Determine what classes the records belong to and which class is most prevalent.

3. Assign that class to the new record.

Alternatively (or in addition), it may be desirable to tweak the method so that it answers the question: “What is the propensity of belonging to the class of interest?” instead of “Which class is the most probable?” Obtaining class prob- abilities allows using a sliding cutoff to classify a record as belonging to class Ci, even if Ci is not the most probable class for that record. This approach is useful when there is a specific class of interest that we are interested in iden- tifying, and we are willing to “overidentify” records as belonging to this class.

187

188 THE NAIVE BAYES CLASSIFIER

(See Chapter 5 for more details on the use of cutoffs for classification and on asymmetric misclassification costs)

Cutoff Probability Method

1. Establish a cutoff probability for the class of interest above which we con- sider that a record belongs to that class.

2. Find all the training records with the same predictor profile as the new record (i.e., where the predictor values are the same).

3. Determine the probability that those records belong to the class of interest.

4. If that probability is above the cutoff probability, assign the new record to the class of interest.

Conditional Probability

Both procedures incorporate the concept of conditional probability, or the prob- ability of event A given that event B has occurred [denoted P (A|B)]. In this case, we will be looking at the probability of the record belonging to class Ci given that its predictor values are x1, x2, . . . , xp. In general, for a response with m classes C1, C2, . . . , Cm, and the predictor values x1, x2, . . . , xp, we want to compute

P (Ci|x1, . . . , xp). (8.1)

To classify a record, we compute its probability of belonging to each of the classes in this way, then classify the record to the class that has the highest probability or use the cutoff probability to decide whether it should be assigned to the class of interest.

From this definition, we see that the Bayesian classifier works only with cat- egorical predictors. If we use a set of numerical predictors, then it is highly unlikely that multiple records will have identical values on these numerical pre- dictors. Therefore, numerical predictors must be binned and converted to cate- gorical predictors. The Bayesian classifier is the only classification or prediction method presented in this book that is especially suited for (and limited to) categorical predictor variables.

Example 1: Predicting Fraudulent Financial Reporting

An accounting firm has many large companies as customers. Each customer submits an annual financial report to the firm, which is then audited by the accounting firm. For simplicity, we will designate the outcome of the audit as “fraudulent” or “truthful,” referring to the accounting firm’s assessment of the customer’s financial report. The accounting firm has a strong incentive to

APPLYING THE FULL (EXACT) BAYESIAN CLASSIFIER 189

be accurate in identifying fraudulent reports—if it passes a fraudulent report as truthful, it would be in legal trouble.

The accounting firm notes that, in addition to all the financial records, it also has information on whether or not the customer has had prior legal trouble (criminal or civil charges of any nature filed against it). This information has not been used in previous audits, but the accounting firm is wondering whether it could be used in the future to identify reports that merit more intensive review. Specifically, it wants to know whether having had prior legal trouble is predictive of fraudulent reporting.

In this case, each customer is a record, and the outcome variable of interest, Y = {fraudulent, truthful}, has two classes into which a company can be classi- fied: C1 = fraudulent and C2 = truthful. The predictor variable—“prior legal trouble”—has two values: 0 (no prior legal trouble) and 1 (prior legal trouble).

The accounting firm has data on 1500 companies that it has investigated in the past. For each company, it has information on whether the financial report was judged fraudulent or truthful and whether the company had prior legal trouble. The data were partitioned into a training set (1000 firms) and a validation set (500 firms). Counts in the training set are shown in Table 8.1.

TABLE 8.1 PIVOT TABLE FOR FINANCIAL REPORTING EXAMPLE

Prior Legal No Prior Legal (X = 1) (X = 0) Total

Fraudulent (C1) 50 50 100

Truthful (C2) 180 720 900

Total 230 770 1000

8.2 Applying the Full (Exact) Bayesian Classifier

Now consider the financial report from a new company, which we wish to clas- sify as either fraudulent or truthful by using these data. To do this, we compute the probabilities, as above, of belonging to each of the two classes.

If the new company had had prior legal trouble, the probability of belonging to the fraudulent class would be P (fraudulent | prior legal) = 50/230 (of the 230 companies with prior legal trouble in the training set, 50 had fraudulent financial reports). The probability of belonging to the other class, “truthful,” is, of course, the remainder = 180/230.

190 THE NAIVE BAYES CLASSIFIER

Using the “Assign to the Most Probable Class” Method

If a company had prior legal trouble, we assign it to the “truthful” class. Similar calculations for the case of no prior legal trouble are left as an exercise to the reader. In this example, using the rule “assign to the most probable class,” all records are assigned to the “truthful” class. This is the same result as the naive rule of “assign all records to the majority class.”

Using the Cutoff Probability Method

In this example, we are more interested in identifying the fraudulent reports— those are the ones that can land the auditor in jail. We recognize that, in order to identify the fraudulent reports, some truthful reports will be misidentified as fraudulent, and the overall classification accuracy may decline. Our approach is, therefore, to establish a cutoff value for the probability of being fraudulent, and classify all records above that value as fraudulent. The Bayesian formula for the calculation of this probability that a record belongs to class Ci is as follows:

P (Ci|x1, . . . , xp) = P (x1, . . . , xp|Ci)P (Ci)

P (x1, . . . , xp|C1)P (C1) + · · ·+ P (x1, . . . , xp|Cm)P (Cm) .

(8.2)

In this example (where frauds are rarer), if the cutoff were established at 0.20, we would classify a prior legal trouble record as fraudulent because P (fraudulent | prior legal) = 50/230 = 0.22. The user can treat this cutoff as a “slider” to be adjusted to optimize performance, like other parameters in any classification model.

Practical Difficulty with the Complete (Exact) Bayes Procedure

The approach outlined above amounts to finding all the records in the sample that are exactly like the new record to be classified in the sense that all the predictor values are all identical. This was easy in the small example presented above, where there was just one predictor.

When the number of predictors gets larger (even to a modest number like 20), many of the records to be classified will be without exact matches. This can be understood in the context of a model to predict voting on the basis of demographic variables. Even a sizable sample may not contain even a single match for a new record who is a male Hispanic with high income from the US Midwest who voted in the last election, did not vote in the prior election, has three daughters and one son, and is divorced. And this is just eight variables, a small number for most data mining exercises. The addition of just a single new variable with five equally frequent categories reduces the probability of a match by a factor of 5.

APPLYING THE FULL (EXACT) BAYESIAN CLASSIFIER 191

Solution: Naive Bayes

In the naive Bayes solution, we no longer restrict the probability calculation to those records that match the record to be classified. Instead we use the entire dataset.

Returning to our original basic classification procedure outlined at the beginning of the chapter, recall that the procedure for classifying a new record was:

1. Find all the other records with the same predictor profile (i.e., where the predictor values are the same).

2. Determine what classes the records belong to and which class is most prevalent.

3. Assign that class to the new record.

The naive Bayes modification (for the basic classification procedure) is as follows:

1. For class C1, estimate the individual conditional probabilities for each predictor P (xj|C1)—these are the probabilities that the predictor value in the record to be classified occurs in class C1. For example, for X1 this probability is estimated by the proportion of x1 values among the C1 records in the training set.

2. Multiply these probabilities by each other, then by the proportion of records belonging to class C1.

3. Repeat Steps 1 and 2 for all the classes.

4. Estimate a probability for class Ci by taking the value calculated in Step 2 for class Ci and dividing it by the sum of such values for all classes.

5. Assign the record to the class with the highest probability for this set of predictor values.

The above steps lead to the naive Bayes formula for calculating the probability that a record with a given set of predictor values x1, . . . , xp belongs to class C1 among m classes. The formula can be written as follows:

Pnb(C1 | x1, . . . xp) =

P (C1)[P (x1 | C1)P (x2 | C1) · · ·P (xp | C1)] P (C1)[P (x1 | C1)P (x2 | C1) · · ·P (xp | C1)] + · · ·+ P (Cm)[P (x1 | Cm)P (x2 | Cm) · · ·P (xp | Cm)]

.

(8.3)

This is a somewhat formidable formula; see Example 2 for a simpler numeri- cal version. Note that all the needed quantities can be obtained from pivot tables of Y vs. each of the categorical predictors.

192 THE NAIVE BAYES CLASSIFIER

The Naive Bayes Assumption of Conditional Independence

In probability terms, we have made a simplifying assumption that the exact con- ditional probability of seeing a record with predictor profile x1, x2, . . . , xp within a certain class, P (x1, x2, . . . , xp|Ci), is well approximated by the product of the individual conditional probabilities P (x1|Ci)× P (x2|Ci) · · · × P (xp|Ci). These two quantities are identical when the predictors are independent within each class.

For example, suppose that “lost money last year” is an additional variable in the accounting fraud example. The simplifying assumption we make with naive Bayes is that, within a given class, we no longer need to look for the records characterized both by “prior legal trouble” and “lost money last year.” Rather, assuming that the two are independent, we can simply multiply the probability of “prior legal trouble” by the probability of “lost money last year.” Of course, complete independence is unlikely in practice, where some correlation between predictors is expected.

In practice, despite the assumption violation, the procedure works quite well—primarily because what is usually needed is not a propensity for each record that is accurate in absolute terms but just a reasonably accurate rank order- ing of propensities. Even when the assumption is violated, the rank ordering of the records’ propensities is typically preserved.

Note that if all we are interested in is a rank ordering, and the denominator remains the same for all classes, it is sufficient to concentrate only on the numer- ator. The disadvantage of this approach is that the probability values it yields (the propensities), while ordered correctly, are not on the same scale as the exact values that the user would anticipate.

Using the Cutoff Probability Method

The above procedure is for the basic case where we seek maximum classification accuracy for all classes. In the case of the relatively rare class of special interest, the procedure is:

1. Establish a cutoff probability for the class of interest above which we con- sider that a record belongs to that class.

2. For the class of interest, compute the probability that each individual predictor value in the record to be classified occurs in the training data.

3. Multiply these probabilities times each other, then times the proportion of records belonging to the class of interest.

4. Estimate the probability for the class of interest by taking the value cal- culated in Step 3 for the class of interest and dividing it by the sum of the similar values for all classes.

APPLYING THE FULL (EXACT) BAYESIAN CLASSIFIER 193

5. If this value falls above the cutoff, assign the new record to the class of interest, otherwise not.

6. Adjust the cutoff value as needed, as a parameter of the model.

Example 2: Predicting Fraudulent Financial Reports, Two Predictors

Let us expand the financial reports example to two predictors, and, using a small subset of data, compare the complete (exact) Bayes calculations to the naive Bayes calculations.

Consider the 10 customers of the accounting firm listed in Table 8.2. For each customer, we have information on whether it had prior legal trouble, whether it is a small or large company, and whether the financial report was found to be fraudulent or truthful. Using this information, we will calculate the conditional probability of fraud, given each of the four possible combinations {y, small}, {y, large}, {n, small}, {n, large}.

TABLE 8.2 INFORMATION ON 10 COMPANIES

Company Prior Legal Trouble Company Size Status

1 Yes Small Truthful 2 No Small Truthful 3 No Large Truthful 4 No Large Truthful 5 No Small Truthful 6 No Small Truthful 7 Yes Small Fraudulent 8 Yes Large Fraudulent 9 No Large Fraudulent 10 Yes Large Fraudulent

Complete (Exact) Bayes Calculations: The probabilities are computed as

P (fraudulent|PriorLegal = y, Size = small) = 1/2 = 0.5

P (fraudulent|PriorLegal = y, Size = large) = 2/2 = 1

P (fraudulent|PriorLegal = n, Size = small) = 0/3 = 0

P (fraudulent|PriorLegal = n, Size = large) = 1/3 = 0.33

Naive Bayes Calculations: Now we compute the naive Bayes probabilities. For the conditional probability of fraudulent behaviors given {PriorLegal = y, Size = small}, the numerator is a multiplication of the proportion of {Prior- Legal = y} instances among the fraudulent companies, times the proportion of {Size = small} instances among the fraudulent companies, times the proportion of fraudulent companies: (3/4)(1/4)(4/10) = 0.075. To get the actual prob- abilities, we must also compute the numerator for the conditional probability

194 THE NAIVE BAYES CLASSIFIER

of truthful behaviors given {PriorLegal = y, Size = small}: (1/6)(4/6)(6/10) = 0.067. The denominator is then the sum of these two conditional probabilities (0.075 + 0.067 = 0.14). The conditional probability of fraudulent behaviors given {PriorLegal = y, Size = small} is therefore 0.075/0.14 = 0.53. In a similar fashion, we compute all four conditional probabilities:

Pnb(fraudulent|PriorLegal = y, Size = small) = (3/4)(1/4)(4/10)

(3/4)(1/4)(4/10) + (1/6)(4/6)(6/10) = 0.53

Pnb(fraudulent|PriorLegal = y, Size = large) = 0.87 Pnb(fraudulent|PriorLegal = n, Size = small) = 0.07 Pnb(fraudulent|PriorLegal = n, Size = large) = 0.31

Note how close these naive Bayes probabilities are to the exact Bayes probabil- ities. Although they are not equal, both would lead to exactly the same classi- fication for a cutoff of 0.5 (and many other values). It is often the case that the rank ordering of probabilities is even closer to the exact Bayes method than the probabilities themselves, and for classification purposes it is the rank orderings that matter.

We now consider a larger numerical example, where information on flights is used to predict flight delays.

Example 3: Predicting Delayed Flights

Predicting flight delays can be useful to a variety of organizations: airport author- ities, airlines, and aviation authorities. At times, joint task forces have been formed to address the problem. If such an organization were to provide ongo- ing real-time assistance with flight delays, it would benefit from some advance notice about flights that are likely to be delayed.

In this simplified illustration, we look at five predictors (see Table 8.3). The outcome of interest is whether or not the flight is delayed (delayed here means arrived more than 15 minutes late). Our data consist of all flights from the Wash- ington, DC area into the New York City area during January 2004. A record is a particular flight. The percentage of delayed flights among these 2201 flights is 19.5%. The data were obtained from the Bureau of Transportation Statistics (available on the web at www.transtats.bts.gov). The goal is to accurately predict whether or not a new flight (not in this dataset), will be delayed. The outcome variable is whether the flight was delayed, and thus it has two classes (1 = delayed and 0 = on time). In addition, information is collected on the predictors listed in Table 8.3.

The data were first partitioned into training (60%) and validation (40%) sets, and then a naive Bayes classifier was applied to the training set (we use package e1071).

APPLYING THE FULL (EXACT) BAYESIAN CLASSIFIER 195

TABLE 8.3 DESCRIPTION OF VARIABLES FOR FLIGHT DELAYS EXAMPLE

Day of Week Coded as 1 = Monday, 2 = Tuesday, ..., 7 = Sunday Sch. Dep. Time Broken down into 18 intervals between 6:00 AM and 10:00 PM Origin Three airport codes: DCA (Reagan National), IAD (Dulles),

BWI (Baltimore–Washington Int’l) Destination Three airport codes: JFK (Kennedy), LGA (LaGuardia), EWR (Newark) Carrier Eight airline codes: CO (Continental), DH (Atlantic Coast), DL (Delta),

MQ (American Eagle), OH (Comair), RU (Continental Express), UA (United), and US (USAirways)

The first part of the output in Table 8.4 shows the ratios of delayed flights and on time flights in the training set (called a priori probabilities), followed by the conditional probabilities for each class, as a function of the predictor values. Note that the conditional probabilities in the naive Bayes output can be replicated simply by using pivot tables on the training data, looking at the proportion of records for each value relative to the entire class. This is illustrated in Table 8.5, which displays the proportion of delayed (or on time) flights by destination airport (each row adds up to 1).

Note that in this example, there are no predictor values that were not rep- resented in the training data.

To classify a new flight, we compute the probability that it will be delayed and the probability that it will be on time. Recall that since both will have the same denominator, we can just compare the numerators. Each numerator is computed by multiplying all the conditional probabilities of the relevant predictor values and, finally, multiplying by the proportion of that class (in this case P̂ (delayed) = 0.197). Let us use an example: to classify a Delta flight from DCA to LGA departing between 10:00 AM and 11:00 AM on a Sunday, we first compute the numerators:

P̂ (delayed|Carrier = DL, Day_Week = 7, Dep_Time = 10, Dest = LGA, Origin = DCA) ∝ (0.115)(0.146)(0.027)(0.400)(0.519)(0.197) = 0.000019 P̂ (ontime|Carrier = DL, Day_Week = 7, Dep_Time = 10, Dest = LGA, Origin = DCA) ∝ (0.198)(0.106)(0.049)(0.537)(0.651)(0.803) = 0.00029

The symbol ∝ means “is proportional to,” reflecting the fact that this calcula- tion deals only with the numerator in the naive Bayes formula (8.3). Compar- ing the numerators, it is therefore, more likely that the flight will be on time. Note that a record with such a combination of predictor values does not exist in the training set, and therefore we use the naive Bayes rather than the exact Bayes.

196 THE NAIVE BAYES CLASSIFIER

TABLE 8.4 NAIVE BAYES CLASSIFIER APPLIED TO FLIGHT DELAYS (TRAINING) DATA

code for running naive Bayes

library(e1071) delays.df <- read.csv("FlightDelays.csv")

# change numerical variables to categorical first delays.df$DAY_WEEK <- factor(delays.df$DAY_WEEK) delays.df$DEP_TIME <- factor(delays.df$DEP_TIME) # create hourly bins departure time delays.df$CRS_DEP_TIME <- factor(round(delays.df$CRS_DEP_TIME/100))

# Create training and validation sets. selected.var <- c(10, 1, 8, 4, 2, 13) train.index <- sample(c(1:dim(delays.df)[1]), dim(delays.df)[1]*0.6) train.df <- delays.df[train.index, selected.var] valid.df <- delays.df[-train.index, selected.var]

# run naive bayes delays.nb <- naiveBayes(Flight.Status ~ ., data = train.df) delays.nb

Partial output

A-priori probabilities: Y delayed ontime

0.197 0.803

Conditional probabilities: DAY_WEEK

Y 1 2 3 4 5 6 7 delayed 0.2154 0.1346 0.1269 0.1269 0.1846 0.0654 0.1462 ontime 0.1311 0.1377 0.1453 0.1755 0.1783 0.1264 0.1057

CRS_DEP_TIME Y 6 7 8 9 10 11 12 13 14 delayed 0.0308 0.0538 0.0692 0.0269 0.0269 0.0154 0.0615 0.0346 0.0462 ontime 0.0632 0.0575 0.0736 0.0557 0.0491 0.0368 0.0613 0.0783 0.0623

CRS_DEP_TIME Y 15 16 17 18 19 20 21

delayed 0.2077 0.0731 0.1346 0.0308 0.0846 0.0192 0.0846 ontime 0.1142 0.0840 0.0981 0.0406 0.0415 0.0283 0.0557

ORIGIN Y BWI DCA IAD

delayed 0.0885 0.5192 0.3923 ontime 0.0632 0.6519 0.2849

DEST Y EWR JFK LGA

delayed 0.381 0.219 0.400 ontime 0.289 0.174 0.538

CARRIER Y CO DH DL MQ OH RU UA US

delayed 0.0692 0.3308 0.1154 0.1731 0.0115 0.2154 0.0154 0.0692 ontime 0.0377 0.2302 0.1981 0.1292 0.0142 0.1840 0.0132 0.1934

APPLYING THE FULL (EXACT) BAYESIAN CLASSIFIER 197

TABLE 8.5 PIVOT TABLE OF FLIGHT STATUS BY DESTINATION AIRPORT (TRAINING DATA)

# use prop.table() with margin = 1 to convert a count table to a proportion table, # where each row sums up to 1 (use margin = 2 for column sums). > prop.table(table(train.df$Flight.Status, train.df$DEST), margin = 1)

EWR JFK LGA delayed 0.381 0.219 0.400 ontime 0.289 0.174 0.538

To compute the actual probability, we divide each of the numerators by their sum:

P̂ (delayed|Carrier =DL, Day_Week= 7, Dep_Time= 10, Dest = LGA, Origin =DCA)=

= 0.000019

0.000019 + 0.00029 =0.06

P̂ (on time|Carrier =DL, Day_Week= 7, Dep_Time= 10, Dest = LGA, Origin =DCA)=

= 0.00029

0.000019 + 0.00029 =0.94

Of course, we rely on software to compute these probabilities for any records of interest (in the training set, the validation set, or for scoring new data). Table 8.6 shows the estimated probability and class for the example flight.

TABLE 8.6 SCORING THE EXAMPLE FLIGHT (PROBABILITY AND CLASS)

code for scoring data using naive Bayes

## predict probabilities pred.prob <- predict(delays.nb, newdata = valid.df, type = "raw") ## predict class membership pred.class <- predict(delays.nb, newdata = valid.df)

df <- data.frame(actual = valid.df$Flight.Status, predicted = pred.class, pred.prob)

df[valid.df$CARRIER == "DL" & valid.df$DAY_WEEK == 7 & valid.df$CRS_DEP_TIME == 10 & valid.df$DEST == "LGA" & valid.df$ORIGIN == "DCA",]

Output

> df[valid.df$CARRIER == "DL" & valid.df$DAY_WEEK == 7 & valid.df$CRS_DEP_TIME == 10 & + valid.df$DEST == "LGA" & valid.df$ORIGIN == "DCA",]

actual predicted delayed ontime 69 ontime ontime 0.0604 0.94 700 ontime ontime 0.0604 0.94

198 THE NAIVE BAYES CLASSIFIER

Finally, to evaluate the performance of the naive Bayes classifier for our data, we use the confusion matrix, lift charts, and all the measures that were described in Chapter 5. For our example, the confusion matrices for the training and validation sets are shown in Table 8.7. We see that the overall accuracy level is around 80% for both the training and validation data. In comparison, a naive rule that would classify all 880 flights in the validation set as on time would have missed the 172 delayed flights, also resulting in a 80% accuracy. Thus, by a simple accuracy measure, the naive Bayes model does no better than the naive rule. However, examining the lift chart (Figure 8.1) shows the strength of the naive Bayes in capturing the delayed flights effectively, when the goal is ranking.

TABLE 8.7 CONFUSION MATRICES FOR FLIGHT DELAY USING A NAIVE BAYES CLASSIFIER

code for confusion matrices

library(caret)

# training pred.class <- predict(delays.nb, newdata = train.df) confusionMatrix(pred.class, train.df$Flight.Status)

# validation pred.class <- predict(delays.nb, newdata = valid.df) confusionMatrix(pred.class, valid.df$Flight.Status)

Partial output

> # training > confusionMatrix(pred.class, train.df$Flight.Status) Confusion Matrix and Statistics

Reference Prediction delayed ontime

delayed 30 39 ontime 230 1021

Accuracy : 0.796 > # validation > confusionMatrix(pred.class, valid.df$Flight.Status) Confusion Matrix and Statistics

Reference Prediction delayed ontime

delayed 21 25 ontime 147 688

Accuracy : 0.805

ADVANTAGES AND SHORTCOMINGS OF THE NAIVE BAYES CLASSIFIER 199

code for creating Figure 8.1

library(gains) gain <- gains(ifelse(valid.df$Flight.Status=="delayed",1,0), pred.prob[,1], groups=100)

plot(c(0,gain$cume.pct.of.total*sum(valid.df$Flight.Status=="delayed"))~c(0,gain$cume.obs), xlab="# cases", ylab="Cumulative", main="", type="l")

lines(c(0,sum(valid.df$Flight.Status=="delayed"))~c(0, dim(valid.df)[1]), lty=2)

0 200 400 600 800

0 50

10 0

15 0

# cases

C um

ul at

iv e

FIGURE 8.1 LIFT CHART OF NAIVE BAYES CLASSIFIER APPLIED TO FLIGHT DELAYS DATA

8.3 Advantages and Shortcomings of the Naive Bayes Classifier

The naive Bayes classifier’s beauty is in its simplicity, computational efficiency, good classification performance, and ability to handle categorical variables directly. In fact, it often outperforms more sophisticated classifiers even when the underlying assumption of independent predictors is far from true. This advantage is especially pronounced when the number of predictors is very large.

Three main issues should be kept in mind, however. First, the naive Bayes classifier requires a very large number of records to obtain good results.

Second, where a predictor category is not present in the training data, naive Bayes assumes that a new record with that category of the predictor has zero probability. This can be a problem if this rare predictor value is important. One example is the binary predictorWeather in the flights delay dataset, which we did not use for analysis, and which denotes bad weather. When the weather was bad, all flights were delayed. Consider another example, where the outcome variable is bought high-value life insurance and a predictor category is owns yacht. If the

200 THE NAIVE BAYES CLASSIFIER

training data have no records with owns yacht = 1, for any new records where owns yacht = 1, naive Bayes will assign a probability of 0 to the outcome variable bought high-value life insurance. With no training records with owns yacht = 1, of course, no data mining technique will be able to incorporate this potentially important variable into the classification model—it will be ignored. With naive Bayes, however, the absence of this predictor actively “outvotes” any other information in the record to assign a 0 to the outcome value (when, in this case, it has a relatively good chance of being a 1). The presence of a large training set (and judicious binning of continuous predictors, if required) helps mitigate this effect. A popular solution in such cases is to replace zero probabilities with non-zero values using a method called smoothing (e.g., Laplace smoothing can be applied by using argument laplace = 0 in function naiveBayes()).

Finally, good performance is obtained when the goal is classification or ranking of records according to their probability of belonging to a certain class. How- ever, when the goal is to estimate the probability of class membership (propensity), this method provides very biased results. For this reason, the naive Bayes method is rarely used in credit scoring (Larsen, 2005).

S P A M F I L T E R I N G

Filtering spam in e-mail has long been a widely familiar application of data min- ing. Spam filtering, which is based in large part on natural language vocabulary, is a natural fit for a naive Bayesian classifier, which uses exclusively categorical variables. Most spam filters are based on this method, which works as follows:

1. Humans review a large number of e-mails, classify them as “spam” or “not spam,” and from these select an equal (also large) number of spam e-mails and non- spam e-mails. This is the training data.

2. These e-mails will contain thousands of words; for each word, compute the fre- quency with which it occurs in the spam dataset, and the frequency with which it occurs in the non-spam dataset. Convert these frequencies into estimated probabilities (i.e., if the word “free” occurs in 500 out of 1000 spam e-mails, and only 100 out of 1000 non-spam e-mails, the probability that a spam e-mail will contain the word “free” is 0.5, and the probability that a non-spam e-mail will contain the word “free” is 0.1).

3. If the only word in a new message that needs to be classified as spam or not spam is “free,” we would classify the message as spam, since the Bayesian posterior probability is 0.5/(0.5+01) or 5/6 that, given the appearance of “free,” the message is spam.

4. Of course, we will have many more words to consider. For each such word, the probabilities described in Step 2 are calculated, and multiplied together, and formula (8.3) is applied to determine the naive Bayes probability of belonging to the classes. In the simple version, class membership (spam or not spam) is determined by the higher probability.

ADVANTAGES AND SHORTCOMINGS OF THE NAIVE BAYES CLASSIFIER 201

5. In a more flexible interpretation, the ratio between the “spam” and “not spam” probabilities is treated as a score for which the operator can establish (and change) a cutoff threshold—anything above that level is classified as spam.

6. Users have the option of building a personalized training database by classify- ing incoming messages as spam or not spam, and adding them to the training database. One person’s spam may be another person’s substance.

It is clear that, even with the “Naive” simplification, this is an enormous com- putational burden. Spam filters now typically operate at two levels—at servers (intercepting some spam that never makes it to your computer) and on individual computers (where you have the option of reviewing it). Spammers have also found ways to ”poison” the vocabulary-based Bayesian approach, by including sequences of randomly selected irrelevant words. Since these words are randomly selected, they are unlikely to be systematically more prevalent in spam than in non-spam, and they dilute the effect of key spam terms such as “Viagra” and “free.” For this rea- son, sophisticated spam classifiers also include variables based on elements other than vocabulary, such as the number of links in the message, the vocabulary in the subject line, determination of whether the “From:” e-mail address is the real originator (anti-spoofing), use of HTML and images, and origination at a dynamic or static IP address (the latter are more expensive and cannot be set up quickly).

202 THE NAIVE BAYES CLASSIFIER

PROBLEMS

8.1 Personal Loan Acceptance. The file UniversalBank.csv contains data on 5000 cus- tomers of Universal Bank. The data include customer demographic information (age, income, etc.), the customer’s relationship with the bank (mortgage, securities account, etc.), and the customer response to the last personal loan campaign (Personal Loan). Among these 5000 customers, only 480 (= 9.6%) accepted the personal loan that was offered to them in the earlier campaign. In this exercise, we focus on two predictors: Online (whether or not the customer is an active user of online banking services) and Credit Card (abbreviated CC below) (does the customer hold a credit card issued by the bank), and the outcome Personal Loan (abbreviated Loan below).

Partition the data into training (60%) and validation (40%) sets.

a. Create a pivot table for the training data with Online as a column variable, CC as a row variable, and Loan as a secondary row variable. The values inside the table should convey the count. In R use functions melt() and cast(), or function table().

b. Consider the task of classifying a customer who owns a bank credit card and is actively using online banking services. Looking at the pivot table, what is the probability that this customer will accept the loan offer? [This is the probability of loan acceptance (Loan = 1) conditional on having a bank credit card (CC = 1) and being an active user of online banking services (Online = 1)].

c. Create two separate pivot tables for the training data. One will have Loan (rows) as a function of Online (columns) and the other will have Loan (rows) as a function of CC.

d. Compute the following quantities [P (A | B) means “the probability of A given B”]: i. P (CC = 1 | Loan = 1) (the proportion of credit card holders among the loan

acceptors)

ii. P (Online = 1 | Loan = 1) iii. P (Loan = 1) (the proportion of loan acceptors)

iv. P (CC = 1 | Loan = 0) v. P (Online = 1 | Loan = 0)

vi. P (Loan = 0)

e. Use the quantities computed above to compute the naive Bayes probability P (Loan = 1 | CC = 1, Online = 1).

f. Compare this value with the one obtained from the pivot table in (b). Which is a more accurate estimate?

g. Which of the entries in this table are needed for computing P (Loan = 1 | CC = 1, Online = 1)? In R, run naive Bayes on the data. Examine the model output on training data, and find the entry that corresponds to P (Loan = 1 | CC = 1, Online = 1). Compare this to the number you obtained in (e).

8.2 Automobile Accidents. The file Accidents.csv contains information on 42,183 actual automobile accidents in 2001 in the United States that involved one of three levels of injury: NO INJURY, INJURY, or FATALITY. For each accident, additional infor- mation is recorded, such as day of week, weather conditions, and road type. A firm might be interested in developing a system for quickly classifying the severity of an accident based on initial reports and associated data in the system (some of which rely on GPS-assisted reporting).

PROBLEMS 203

Our goal here is to predict whether an accident just reported will involve an injury (MAX_SEV_IR = 1 or 2) or will not (MAX_SEV_IR = 0). For this purpose, create a dummy variable called INJURY that takes the value “yes” if MAX_SEV_IR = 1 or 2, and otherwise “no.”

a. Using the information in this dataset, if an accident has just been reported and no further information is available, what should the prediction be? (INJURY = Yes or No?) Why?

b. Select the first 12 records in the dataset and look only at the response (INJURY) and the two predictors WEATHER_R and TRAF_CON_R.

i. Create a pivot table that examines INJURY as a function of the two predictors for these 12 records. Use all three variables in the pivot table as rows/columns.

ii. Compute the exact Bayes conditional probabilities of an injury (INJURY = Yes) given the six possible combinations of the predictors.

iii. Classify the 12 accidents using these probabilities and a cutoff of 0.5.

iv. Compute manually the naive Bayes conditional probability of an injury given WEATHER_R = 1 and TRAF_CON_R = 1.

v. Run a naive Bayes classifier on the 12 records and two predictors using R. Check the model output to obtain probabilities and classifications for all 12 records. Compare this to the exact Bayes classification. Are the resulting clas- sifications equivalent? Is the ranking (= ordering) of observations equivalent?

c. Let us now return to the entire dataset. Partition the data into training (60%) and validation (40%).

i. Assuming that no information or initial reports about the accident itself are available at the time of prediction (only location characteristics, weather con- ditions, etc.), which predictors can we include in the analysis? (Use the Data_Codes sheet.)

ii. Run a naive Bayes classifier on the complete training set with the relevant pre- dictors (and INJURY as the response). Note that all predictors are categorical. Show the confusion matrix.

iii. What is the overall error for the validation set?

iv. What is the percent improvement relative to the naive rule (using the validation set)?

v. Examine the conditional probabilities output. Why do we get a probability of zero for P(INJURY = No | SPD_LIM = 5)?

CHAPTER 9

Classification and Regression Trees

This chapter describes a flexible data-driven method that can be used for both classification (called classification tree) and prediction (called regression tree). Among the data-driven methods, trees are the most transparent and easy to interpret. Trees are based on separating records into subgroups by creating splits on pre- dictors. These splits create logical rules that are transparent and easily under- standable, for example, “IF Age < 55 AND Education > 12 THEN class = 1.” The resulting subgroups should be more homogeneous in terms of the outcome variable, thereby creating useful prediction or classification rules. We discuss the two key ideas underlying trees: recursive partitioning (for constructing the tree) and pruning (for cutting the tree back). In the context of tree construction, we also describe a few metrics of homogeneity that are popular in tree algorithms, for determining the homogeneity of the resulting subgroups of records. We explain that pruning is a useful strategy for avoiding overfitting and show how it is done. We also describe alternative strategies for avoiding overfitting. As with other data-driven methods, trees require large amounts of data. However, once con- structed, they are computationally cheap to deploy even on large samples. They also have other advantages such as being highly automated, robust to outliers, and able to handle missing values. In addition to prediction and classification, we describe how trees can be used for dimension reduction. Finally, we intro- duce random forests and boosted trees, which combine results from multiple trees to improve predictive power.

9.1 Introduction

If one had to choose a classification technique that performs well across a wide range of situations without requiring much effort from the analyst while being

205

206 CLASSIFICATION AND REGRESSION TREES

Income < 106

Education < 1.5

Family < 2.5

Income < 114

Income < 116

CCAvg < 2.4

2712 288

2312 31

400 257

366 45

359 0

7 45

7 4 0 41

34 212

34 22

29 7 5 15

0 190

0

0

0

0

0

1

0 1

1

0

0 1

1

yes no

FIGURE 9.1 BEST-PRUNED TREE OBTAINED BY FITTING A FULL TREE TO THE TRAINING DATA

readily understandable by the consumer of the analysis, a strong contender would be the tree methodology developed by Breiman et al. (1984). We discuss this classification procedure first, then in later sections we show how the proce- dure can be extended to prediction of a numerical outcome. The program that Breiman et al. created to implement these procedures was called CART (Clas- sification And Regression Trees). A related procedure is called C4.5.

What is a classification tree? Figure 9.1 shows a tree for classifying bank customers who receive a loan offer as either acceptors or nonacceptors, based on information such as their income, education level, and average credit card expenditure. One of the reasons that tree classifiers are very popular is that they provide easily understandable classification rules (at least if the trees are not too large). Consider the tree in the example. The gray terminal nodes are marked with 0 or 1 corresponding to a nonacceptor (0) or acceptor (1). The values above the white nodes give the splitting value on a predictor. The values below the nodes gives the number of records in the split. This tree can easily be translated into a set of rules for classifying a bank customer. For example, the bottom-left rectangle node under the “Family” circle in this tree gives us the following rule:

IF(Income ≥ 106) AND (Education < 1.5) AND (Family < 2.5) THEN Class = 0 (nonacceptor).

In the following, we show how trees are constructed and evaluated.

CLASSIFICATION TREES 207

9.2 Classification Trees

Two key ideas underlie classification trees. The first is the idea of recursive parti- tioning of the space of the predictor variables. The second is the idea of pruning using validation data. In the next few sections, we describe recursive partitioning and in subsequent sections explain the pruning methodology.

Recursive Partitioning

Let us denote the outcome variable by Y and the input (predictor) variables by X1, X2, X3, . . . , Xp. In classification, the outcome variable will be a categori- cal variable. Recursive partitioning divides up the p-dimensional space of the X predictor variables into nonoverlapping multidimensional rectangles. The pre- dictor variables here are considered to be continuous, binary, or ordinal. This division is accomplished recursively (i.e., operating on the results of prior divi- sions). First, one of the predictor variables is selected, say Xi, and a value of Xi, say si, is chosen to split the p-dimensional space into two parts: one part that contains all the points with Xi < si and the other with all the points with Xi ≥ si. Then, one of these two parts is divided in a similar manner by again choosing a predictor variable (it could be Xi or another variable) and a split value for that variable. This results in three (multidimensional) rectangular regions. This process is continued so that we get smaller and smaller rectangu- lar regions. The idea is to divide the entire X-space up into rectangles such that each rectangle is as homogeneous or “pure” as possible. By pure, we mean containing records that belong to just one class. (Of course, this is not always possible, as there may be records that belong to different classes but have exactly the same values for every one of the predictor variables.)

Let us illustrate recursive partitioning with an example.

Example 1: Riding Mowers

We again use the riding-mower example presented in Chapter 3. A riding- mower manufacturer would like to find a way of classifying families in a city into those likely to purchase a riding mower and those not likely to buy one. A pilot random sample of 12 owners and 12 nonowners in the city is undertaken. The data are shown and plotted in Table 9.1 and Figure 9.2.

If we apply the classification tree procedure to these data, the procedure will choose Income for the first split with a splitting value of 60. The (X1, X2) space is now divided into two rectangles, one with Income < 60 and the other with Income ≥ 60. This is illustrated in Figure 9.3.

Notice how the split has created two rectangles, each of which is much more homogeneous than the rectangle before the split. The left rectangle contains

208 CLASSIFICATION AND REGRESSION TREES

TABLE 9.1 LOT SIZE, INCOME, AND OWNERSHIP OF A RIDING MOWER FOR 24 HOUSEHOLDS

Household Income Lot Size Ownership of Number ($000s) (000s ft2) Riding Mower

1 60.0 18.4 Owner 2 85.5 16.8 Owner 3 64.8 21.6 Owner 4 61.5 20.8 Owner 5 87.0 23.6 Owner 6 110.1 19.2 Owner 7 108.0 17.6 Owner 8 82.8 22.4 Owner 9 69.0 20.0 Owner

10 93.0 20.8 Owner 11 51.0 22.0 Owner 12 81.0 20.0 Owner 13 75.0 19.6 Nonowner 14 52.8 20.8 Nonowner 15 64.8 17.2 Nonowner 16 43.2 20.4 Nonowner 17 84.0 17.6 Nonowner 18 49.2 17.6 Nonowner 19 59.4 16.0 Nonowner 20 66.0 18.4 Nonowner 21 47.4 16.4 Nonowner 22 33.0 18.8 Nonowner 23 51.0 14.0 Nonowner 24 63.0 14.8 Nonowner

FIGURE 9.2 SCATTER PLOT OF LOT SIZE VS. INCOME FOR 24 OWNERS AND NONOWNERS OF RIDING MOWERS

CLASSIFICATION TREES 209

FIGURE 9.3 SPLITTING THE 24 RECORDS BY INCOME VALUE OF 60

points that are mostly nonowners (seven nonowners and one owner) and the right rectangle contains mostly owners (11 owners and five nonowners).

How was this particular split selected? The algorithm examined each pre- dictor variable (in this case, Income and Lot Size) and all possible split values for each variable to find the best split. What are the possible split values for a variable? They are simply the values for each predictor. The possible split points for Income are {33.0, 43.2, 47.4, . . . , 110.1} and those for Lot Size are {14.0, 14.8, 16.0, . . . , 23.6}. These split points are ranked according to how much they reduce impurity (heterogeneity) in the resulting rectangle. A pure rectangle is one that is composed of a single class (e.g., owners). The reduction in impurity is defined as overall impurity before the split minus the sum of the impurities for the two rectangles that result from a split.

Categorical Predictors The previous description used numerical predic- tors; however, categorical predictors can also be used in the recursive partitioning context. To handle categorical predictors, the split choices for a categorical pre- dictor are all ways in which the set of categories can be divided into two subsets. For example, a categorical variable with four categories, say {a, b, c, d}, can be split in seven ways into two subsets: {a} and {b, c, d}; {b} and {a, c, d}; {c} and {a, b, d}; {d} and {a, b, c}; {a, b} and {c, d}; {a, c} and {b, d}; and finally {a, d} and {b, c}. When the number of categories is large, the number of splits becomes very large.

210 CLASSIFICATION AND REGRESSION TREES

Measures of Impurity

There are a number of ways to measure impurity. The two most popular mea- sures are the Gini index and an entropy measure. We describe both next. Denote the m classes of the response variable by k = 1, 2, . . . ,m.

The Gini impurity index for a rectangle A is defined by

I(A) = 1− m∑ k=1

p2k,

where pk is the proportion of records in rectangle A that belong to class k. This measure takes values between 0 (when all the records belong to the same class) and (m− 1)/m (when all m classes are equally represented). Figure 9.4 shows the values of the Gini index for a two-class case as a function of pk. It can be seen that the impurity measure is at its peak when pk = 0.5 (i.e., when the rectangle contains 50% of each of the two classes).

FIGURE 9.4 VALUES OF THE GINI INDEX FOR A TWO-CLASS CASE AS A FUNCTION OF THE PROPORTION OF RECORDS IN CLASS 1 (p1)

A second impurity measure is the entropy measure. The entropy for a rect- angle A is defined by

entropy(A) = − m∑ k=1

pk log2(pk)

(to compute log2(x) in R, use function log2()). This measure ranges between 0 (most pure, all records belong to the same class) and log2(m) (when all m classes are represented equally). In the two-class case, the entropy measure is maximized (like the Gini index) at pk = 0.5.

Let us compute the impurity in the riding mower example before and after the first split (using Income with the value of 60). The unsplit dataset contains

CLASSIFICATION TREES 211

12 owners and 12 nonowners. This is a two-class case with an equal number of records from each class. Both impurity measures are therefore at their maximum value: Gini = 0.5 and entropy = log2(2) = 1. After the split, the left rectan- gle contains seven nonowners and one owner. The impurity measures for this rectangle are:

Gini_left = 1− (7/8)2 − (1/8)2 = 0.219

entropy_left = −(7/8) log2(7/8)− (1/8) log2(1/8) = 0.544

The right rectangle contains 11 owners and five nonowners. The impurity mea- sures of the right rectangle are therefore

Gini_right = 1− (11/16)2 − (5/16)2 = 0.430

entropy_right = −(11/16) log2(11/16)− (5/16) log2(5/16) = 0.896

The combined impurity of the two rectangles that were created by the split is a weighted average of the two impurity measures, weighted by the number of records in each:

Gini = (8/24)(0.219) + (16/24)(0.430) = 0.359

entropy = (8/24)(0.544) + (16/24)(0.896) = 0.779

Thus, the Gini impurity index decreased from 0.5 before the split to 0.359 after the split. Similarly, the entropy impurity measure decreased from 1 before the split to 0.779 after the split.

By comparing the reduction in impurity across all possible splits in all possible predictors, the next split is chosen. If we continue splitting the mower data, the next split is on the Lot Size variable at the value 21. Figure 9.5 shows that once again the tree procedure has astutely chosen to split a rectangle to increase the purity of the resulting rectangles. The lower-left rectangle, which contains data points with Income < 60 and Lot Size < 21, has all points that are nonowners; whereas the upper left rectangle, which contains data points with Income < 60 and Lot Size ≥ 21, consists exclusively of a single owner. In other words, the two left rectangles are now “pure.” We can see how the recursive partitioning is refining the set of constituent rectangles to become purer as the algorithm proceeds. The final stage of the recursive partitioning is shown in Figure 9.6. Notice that each rectangle is now pure: it contains data points from just one of the two classes.

The reason the method is called a classification tree algorithm is that each split can be depicted as a split of a node into two successor nodes. The first split is shown as a branching of the root node of a tree in Figure 9.7. The full-grown tree is shown in Figure 9.8. (Note that in R the split values are integers).

212 CLASSIFICATION AND REGRESSION TREES

FIGURE 9.5 SPLITTING THE 24 RECORDS FIRST BY INCOME VALUE OF 60 AND THEN LOT SIZE VALUE OF 21

FIGURE 9.6 FINAL STAGE OF RECURSIVE PARTITIONING; EACH RECTANGLE CONSISTING OF A SINGLE CLASS (OWNERS OR NONOWNERS)

CLASSIFICATION TREES 213

code for running and plotting classification tree with single split

library(rpart) library(rpart.plot) mower.df <- read.csv("RidingMowers.csv")

# use rpart() to run a classification tree. # define rpart.control() in rpart() to determine the depth of the tree. class.tree <- rpart(Ownership ~ ., data = mower.df,

control = rpart.control(maxdepth = 2), method = "class") ## plot tree # use prp() to plot the tree. You can control plotting parameters such as color, shape, # and information displayed (which and where). prp(class.tree, type = 1, extra = 1, split.font = 1, varlen = -10)

Income < 60

12 12

7 1 5 11

Nonowner

Nonowner Owner

yes no

FIGURE 9.7 TREE REPRESENTATION OF FIRST SPLIT (CORRESPONDS TO FIGURE 9.3)

Income < 60

Lot_Size < 21 Lot_Size < 20

Income < 85

Income >= 62

12 12

7 1

7 0 0 1

5 11

5 4

5 1

5 0 0 1

0 3

0 7

Nonowner

Nonowner

Nonowner Owner

Owner

Nonowner

Nonowner

Nonowner Owner

Owner

Owner

yes no

FIGURE 9.8 TREE REPRESENTATION AFTER ALL SPLITS (CORRESPONDS TO FIGURE 9.6). THIS IS THE FULL GROWN TREE

214 CLASSIFICATION AND REGRESSION TREES

Tree Structure

We have two types of nodes in a tree: decision nodes and terminal nodes. Nodes that have successors are called decision nodes because if we were to use a tree to classify a new record for which we knew only the values of the predictor vari- ables, we would “drop” the record down the tree so that at each decision node, the appropriate branch is taken until we get to a node that has no successors. Such nodes are called the terminal nodes (or leaves of the tree), and represent the partitioning of the data by predictors.

It is useful to note that the type of trees grown by R’s rpart() function, also known as CART or binary trees, have the property that the number of terminal nodes is exactly one more than the number of decision nodes.

When using the R function prp() for plotting a tree, decision nodes are depicted by white ovals, and terminal nodes by gray ovals.

The name of the variable chosen for splitting and its splitting value are above each decision node. The numbers below a node are the number of records in that node that had values lesser than (left side) or larger or equal to (right side) the splitting value. (“yes” and “no” tags at the top node clarify on which side we find the lesser values).

Classifying a New Record

To classify a new record, it is “dropped” down the tree. When it has dropped all the way down to a terminal node, we can assign its class simply by taking a “vote” of all the training data that belonged to the terminal node when the tree was grown. The class with the highest vote is assigned to the new record. For instance, a new record reaching the rightmost terminal node in Figure 9.8, which has a majority of records that belong to the owner class, would be classified as “owner.” Alternatively, if a single class is of interest, the algorithm counts the number of “votes” for this class, converts it to a proportion (propensity), then compares it to a user-specified cutoff value. See Chapter 5 for further discussion of the use of a cutoff value in classification, for cases where a single class is of interest.

In a binary classification situation (typically, with a success class that is rel- atively rare and of particular interest), we can also establish a lower cutoff to better capture those rare successes (at the cost of lumping in more failures as suc- cesses). With a lower cutoff, the votes for the success class only need attain that lower cutoff level for the entire terminal node to be classified as a success. The cutoff therefore determines the proportion of votes needed for determining the terminal node class.

EVALUATING THE PERFORMANCE OF A CLASSIFICATION TREE 215

9.3 Evaluating the Performance of a Classification Tree

We have seen with previous methods that the modeling job is not completed by fitting a model to training data; we need out-of-sample data to assess and tune the model. This is particularly true with classification and regression trees, for two reasons:

• Tree structure can be quite unstable, shifting substantially depending on the sample chosen.

• A fully-fit tree will invariably lead to overfitting.

To visualize the first challenge, potential instability, imagine that we partition the data randomly into two samples, A and B, and we build a tree with each. If there are several predictors of roughly equal predictive power, you can see that it would be easy for samples A and B to select different predictors for the top level split, just based on which records ended up in which sample. And a different split at the top level would likely cascade down and yield completely different sets of rules. So we should view the results of a single tree with some caution.

To illustrate the second challenge, overfitting, let’s examine another example.

Example 2: Acceptance of Personal Loan

Universal Bank is a relatively young bank that is growing rapidly in terms of overall customer acquisition. The majority of these customers are liability cus- tomers with varying sizes of relationship with the bank. The customer base of asset customers is quite small, and the bank is interested in growing this base rapidly to bring in more loan business. In particular, it wants to explore ways of converting its liability (deposit) customers to personal loan customers.

A campaign the bank ran for liability customers showed a healthy conversion rate of over 9% successes. This has encouraged the retail marketing department to devise smarter campaigns with better target marketing. The goal of our analysis is to model the previous campaign’s customer behavior to analyze what combi- nation of factors make a customer more likely to accept a personal loan. This will serve as the basis for the design of a new campaign.

Our predictive model will be a classification tree. To assess the accuracy of the tree in classifying new records, we start with the tools and criteria discussed in Chapter 5—partitioning the data into training and validation sets, and later introduce the idea of cross-validation.

The bank’s dataset includes data on 5000 customers. The data include cus- tomer demographic information (age, income, etc.), customer response to the last personal loan campaign (Personal Loan), and the customer’s relationship with

216 CLASSIFICATION AND REGRESSION TREES

the bank (mortgage, securities account, etc.). Table 9.2 shows a sample of the bank’s customer database for 20 customers, to illustrate the structure of the data. Among these 5000 customers, only 480 (= 9.6%) accepted the personal loan that was offered to them in the earlier campaign.

After randomly partitioning the data into training (3000 records) and vali- dation (2000 records), we use the training data to construct a tree. A tree with 7 splits is shown in Figure 9.9 (this is the default tree produced by rpart() for this data). The top node refers to all the records in the training set, of which 2709 customers did not accept the loan and 291 customers accepted the loan. The “0” in the top node’s rectangle represents the majority class (“did not accept” = 0). The first split, which is on the income variable, generates left and right child nodes. To the left is the child node with customers who have income less than 114. Customers with income greater than or equal to 114 go to the right. The splitting process continues; where it stops depends on the parameter settings of the algorithm. The eventual classification of customer appears in the terminal nodes. Of the eight terminal nodes, four lead to classification of “did not accept” and four lead to classification of “accept.”

A full-grown tree is shown in Figure 9.10. Although it is difficult to see the exact splits, let us assess the performance of this tree with the validation data and compare it to the smaller default tree. Each record in the validation data is “dropped down” the tree and classified according to the terminal node it reaches. These predicted classes can then be compared to the actual memberships via a confusion matrix. When a particular class is of interest, a lift chart is useful for assessing the model’s ability to capture those members.

Table 9.3 displays the confusion matrices for the training and validation sets of the small default tree (top) and for the full tree (bottom). Comparing the two training matrices, we see that the full tree has higher accuracy: it is 100% accurate in classifying the training data, which means it has completely pure terminal nodes. In contrast, the confusion matrices for the validation data show that the smaller tree is more accurate. The main reason is that the full-grown tree overfits the training data (to perfect accuracy!). This motivates the next section, where we describe ways to avoid overfitting by either stopping the growth of the tree before it is fully grown or by pruning the full-grown tree.

9.4 Avoiding Overfitting

One danger in growing deeper trees on the training data is overfitting. As dis- cussed in Chapter 5, overfitting will lead to poor performance on new data. If we look at the overall error at the various sizes of the tree, it is expected to decrease as the number of terminal nodes grows until the point of overfitting. Of course, for the training data the overall error decreases more and more until it is zero

AVOIDING OVERFITTING 217

T A B L E

9 .2

SA M PL

E OF

DA TA

FO R 20

CU ST OM

ER S OF

U N IV ER

SA L BA

N K

Pr of es si on

al Fa m ily

CC Pe rs on

al Se cu ri ti es

CD On

lin e

Cr ed

it ID

Ag e

Ex pe

ri en

ce In co m e

Si ze

Av g

Ed uc at io n

M or tg ag e

Lo an

Ac co un

t Ac

co un

t Ba

nk in g

Ca rd

1 25

1 49

4 1. 60

U G

0 No

Ye s

No No

No 2

45 19

34 3

1. 50

U G

0 No

Ye s

No No

No 3

39 15

11 1

1. 00

U G

0 No

No No

No No

4 35

9 10

0 1

2. 70

Gr ad

0 No

No No

No No

5 35

8 45

4 1. 00

Gr ad

0 No

No No

No Ye

s 6

37 13

29 4

0. 40

Gr ad

15 5

No No

No Ye

s No

7 53

27 72

2 1. 50

Gr ad

0 No

No No

Ye s

No 8

50 24

22 1

0. 30

Pr of

0 No

No No

No Ye

s 9

35 10

81 3

0. 60

Gr ad

10 4

No No

No Ye

s No

10 34

9 18

0 1

8. 90

Pr of

0 Ye

s No

No No

No 11

65 39

10 5

4 2. 40

Pr of

0 No

No No

No No

12 29

5 45

3 0. 10

Gr ad

0 No

No No

Ye s

No 13

48 23

11 4

2 3. 80

Pr of

0 No

Ye s

No No

No 14

59 32

40 4

2. 50

Gr ad

0 No

No No

Ye s

No 15

67 41

11 2

1 2. 00

U G

0 No

Ye s

No No

No 16

60 30

22 1

1. 50

Pr of

0 No

No No

Ye s

Ye s

17 38

14 13

0 4

4. 70

Pr of

13 4

Ye s

No No

No No

18 42

18 81

4 2. 40

U G

0 No

No No

No No

19 46

21 19

3 2

8. 10

Pr of

0 Ye

s No

No No

No 20

55 28

21 1

0. 50

Gr ad

0 No

Ye s

No No

Ye s

218 CLASSIFICATION AND REGRESSION TREES

code for creating a default classification tree

library(rpart) library(rpart.plot)

bank.df <- read.csv("UniversalBank.csv") bank.df <- bank.df[ , -c(1, 5)] # Drop ID and zip code columns.

# partition set.seed(1) train.index <- sample(c(1:dim(bank.df)[1]), dim(bank.df)[1]*0.6) train.df <- bank.df[train.index, ] valid.df <- bank.df[-train.index, ]

# classification tree default.ct <- rpart(Personal.Loan ~ ., data = train.df, method = "class") # plot tree prp(default.ct, type = 1, extra = 1, under = TRUE, split.font = 1, varlen = -10)

Income < 114

CCAvg < 3

CD.Account < 0.5

Income < 92

Family < 2.5

Education < 1.5

Family < 2.5

2709 291

2360 50

2233 9

127 41

123 31

89 10

34 21

32 7 2 14

4 10

349 241

342 39

342 0 0 39

7 202

0

0

0

0

0

0

0

0 1

1

0

0

0 1

1

yes no

FIGURE 9.9 DEFAULT CLASSIFICATION TREE FOR THE LOAN ACCEPTANCE DATA USING THE TRAINING SET (3000 RECORDS)

AVOIDING OVERFITTING 219

In co

m e

< 1

14

C C

A vg

< 3

In co

m e

< 1

06

Fa m

ily <

3 .5

E xp

er ie

nc e

> =

1 6

A ge

< 4

2

In co

m e

> =

1 08

Fa m

ily <

2 .5

A ge

> =

2 8

A ge

< 2

8

A ge

< 3

3

C C

A vg

> =

1 .1

E du

ca tio

n >

= 1

.5

In co

m e

> =

1 08

A ge

< 4

2

C D

.A cc

ou nt

< 0

.5

In co

m e

< 9

2

A ge

> =

3 0

E du

ca tio

n >

= 2

.5

Fa m

ily >

= 2

.5

C C

A vg

> =

3 .2

C C

A vg

< 3

C C

A vg

> =

3 .2

A ge

< 4

4

A ge

> =

3 4

A ge

< 3

2

M or

tg ag

e >

= 1

09

A ge

> =

4 4

Fa m

ily >

= 1

.5

A ge

> =

6 1

In co

m e

> =

9 2

Fa m

ily <

2 .5

E du

ca tio

n <

1 .5

A ge

< 5

6

E xp

er ie

nc e

> =

3 4

In co

m e

< 1

04

A ge

< 4

6

Fa m

ily <

1 .5

In co

m e

> =

1 10

A ge

> =

6 0

A ge

< 6

3

C C

A vg

> =

4

M or

tg ag

e <

1 96

In co

m e

> =

9 2

A ge

< 4

2

E du

ca tio

n <

1 .5

Fa m

ily <

2 .5

In co

m e

< 1

16

C C

A vg

< 2

.2

O nl

in e

> =

0 .5

A ge

> =

5 9

M or

tg ag

e >

= 1

42

27 09

2 91

23 60

5 0

22 33

9

21 83

0

50 9

39 4

23 0

16 4

16 3

16 2

15 1

12 0

3 1

3 0

0 1

1 1

1 0

0 1

0 1

0 1

11 5

11 2

11 1

10 0

1 1

1 0

0 10 1

0 3

12 7

4 1

12 3

3 1

89 1

0

88 7

38 0

50 7

30 2

21 0

9 2

9 0

0 2

20 5

20 4

13 1

12 0

1 1

1 0

0 1

7 3

4 0

3 3

3 2

2 0

1 2

1 0

0 20 1

0 1

1 3

1 0

0 3

34 2

1

32 7

26 1

25 0

1 1

1 0

0 1

6 6

6 2

5 0

1 2

1 0

0 20 4

2 1

4

1 0

1 1

4

1 1

1 0

0 1

0 1

3

4 1

0

4 4

4 1

3 0

1 1

1 0

0 10 3

0 6

34 9

2 41

34 2

3 9

34 2

0 0

3 9

7 2

02

7 1

0

6 2

5 0

1 2

1 0

0 2

1 8

1 0

0 8

0 1

92

0

0

0

0

0

0

0

0

0

0

0

0

0

0 1

0

0 1

1

1

0

0

0

0

0

0 11

1

0

0

0

0

0

0

0

0

0

0 1

0

0

0

0

0

0 1

0

0

0

0

0

1

0 11

1

1

0 1

0

0

0

0

0

0 1

0

0

0

1

0 11

1

0

1

0

0 1

1

1

0

0

0

0

0 11

1

0

0

0 1

1

1

0

0

1

0 1

1

0 1

1

ye s

no

F IG

U R E

9 .1

0 A FU

LL TR EE

FO R TH

E LO

AN AC

CE PT AN

CE DA

TA U SI N G TH

E TR AI N IN

G SE T (3

00 0 RE

CO RD

S)

code for creating a deeper classification tree

deeper.ct <- rpart(Personal.Loan ~ ., data = train.df, method = "class", cp = 0, minsplit = 1) # count number of leaves length(deeper.ct$frame$var[deeper.ct$frame$var == "<leaf>"]) # plot tree prp(deeper.ct, type = 1, extra = 1, under = TRUE, split.font = 1, varlen = -10,

box.col=ifelse(deeper.ct$frame$var == "<leaf>", 'gray', 'white'))

220 CLASSIFICATION AND REGRESSION TREES

TABLE 9.3 CONFUSION MATRICES AND ACCURACY FOR THE DEFAULT (SMALL) AND DEEPER (FULL) CLASSIFICATION TREES, ON THE TRAINING AND VALIDATION SETS OF THE PERSONAL LOAN DATA

code for classifying the validation data using a tree and computing the confusion matrices and accuracy for the training and validation data

# classify records in the validation data. # set argument type = "class" in predict() to generate predicted class membership. default.ct.point.pred.train <- predict(default.ct,train.df,type = "class") # generate confusion matrix for training data confusionMatrix(default.ct.point.pred.train, train.df$PersonalLoan) ### repeat the code for the validation set, and the deeper tree

Output

> # default tree: training > confusionMatrix(default.ct.point.pred.train, train.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 2696 26 1 13 265

Accuracy : 0.987

> # default tree: validation > confusionMatrix(default.ct.point.pred.valid, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1792 18 1 19 171

Accuracy : 0.9815

> # deeper tree: training > confusionMatrix(deeper.ct.point.pred.train, train.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 2709 0 1 0 291

Accuracy : 1

> # deeper tree: validation > confusionMatrix(deeper.ct.point.pred.valid, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1788 19 1 23 170

Accuracy : 0.979

AVOIDING OVERFITTING 221

FIGURE 9.11 ERROR RATE AS A FUNCTION OF THE NUMBER OF SPLITS FOR TRAINING VS. VALIDATION DATA: OVERFITTING

at the maximum level of the tree. However, for new data, the overall error is expected to decrease until the point where the tree fully models the relationship between class and the predictors. After that, the tree starts to model the noise in the training set, and we expect the overall error for the validation set to start increasing. This is depicted in Figure 9.11. One intuitive reason a large tree may overfit is that its final splits are based on very small numbers of records. In such cases, class difference is likely to be attributed to noise rather than predictor information.

Stopping Tree Growth: Conditional Inference Trees

One can think of different criteria for stopping the tree growth before it starts overfitting the data. Examples are tree depth (i.e., number of splits), minimum number of records in a terminal node, and minimum reduction in impurity. In R’s rpart(), for example, we can control the depth of the tree with the complexity parameter (CP). The problem is that it is not simple to determine what is a good stopping point using such rules.

Previous methods developed were based on the idea of recursive partition- ing, using rules to prevent the tree from growing excessively and overfitting the training data. One popular method called CHAID (chi-squared automatic interaction detection) is a recursive partitioning method that predates classifica- tion and regression tree (CART) procedures by several years and is widely used in database marketing applications to this day. It uses a well-known statistical test (the chi-square test for independence) to assess whether splitting a node improves the purity by a statistically significant amount. In particular, at each node, we split on the predictor with the strongest association with the outcome variable. The strength of association is measured by the p-value of a chi-squared test of independence. If for the best predictor the test does not show a sig- nificant improvement, the split is not carried out, and the tree is terminated.

222 CLASSIFICATION AND REGRESSION TREES

This method is more suitable for categorical predictors, but it can be adapted to continuous predictors by binning the continuous values into categorical bins.

A more general class of trees based on this idea is called conditional inference trees (see Hothorn et al., 2006). The R implementation is given in the party and partykit packages, and is suitable for both numerical and categorical outcome and predictor variables.

Pruning the Tree

An alternative popular solution that has proven to be more successful than stop- ping tree growth is pruning the full-grown tree. This is the basis of methods such as CART (developed by Breiman et al., implemented in multiple data mining software packages such as R’s rpart package, SAS Enterprise Miner, CART, and MARS) and C4.5 (developed by Quinlan and implemented in packages such as IBM SPSS Modeler). In C4.5, the training data are used both for growing and pruning the tree. In CART, the innovation is to use the validation data to prune back the tree that is grown from training data. CART and CART-like procedures use validation data to prune back the tree that has deliberately been overgrown using the training data.

The idea behind pruning is to recognize that a very large tree is likely to overfit the training data, and that the weakest branches, which hardly reduce the error rate, should be removed. In the mower example, the last few splits resulted in rectangles with very few points (four rectangles in the full tree had a single record). We can see intuitively that these last splits are likely just capturing noise in the training set rather than reflecting patterns that would occur in future data, such as the validation data. Pruning consists of successively selecting a decision node and redesignating it as a terminal node [lopping off the branches extending beyond that decision node (its subtree) and thereby reducing the size of the tree]. The pruning process trades off misclassification error in the validation dataset against the number of decision nodes in the pruned tree to arrive at a tree that captures the patterns—but not the noise—in the training data.

Cross-Validation

Pruning the tree with the validation data solves the problem of overfitting, but it does not address the problem of instability. Recall that the CART algorithm may be unstable in choosing one or another variable for the top-level splits, and this effect then cascades down and produces highly variable rule sets. The solution is to avoid relying on just one partition of the data into training and validation. Rather, we do so repeatedly using cross-validation (see below), then pool the results. Of course, just accumulating a set of different trees with their different rules will not do much by itself. However, we can use the results from all those trees to learn how deep to grow the original tree. In this process, we introduce

AVOIDING OVERFITTING 223

a parameter that can measure, and control, how deep we grow the tree. We will note this parameter value for each minimum-error tree in the cross-validation process, take an average, then apply that average to limit tree growth to this optimal depth when working with new data.

The cost complexity (CC) of a tree is equal to its misclassification error (based on the training data) plus a penalty factor for the size of the tree. For a tree T that has L(T ) terminal nodes, the cost complexity can be written as

CC(T ) = err(T ) + αL(T ),

where err(T ) is the fraction of training records that are misclassified by tree T and α is a penalty factor for tree size. When α = 0, there is no penalty for having too many nodes in a tree, and this yields a tree using the cost complexity criterion that is the full-grown unpruned tree. When we increase α to a very large value the penalty cost component swamps the misclassification error component of the cost complexity criterion, and the result is simply the tree with the fewest terminal nodes: namely, the tree with one node. So there is a range of trees, from tiny to large, corresponding to a range of α, from large to small.

Returning to the cross-validation process, we can now associate a value of α with the minimum error tree developed in each iteration of that process.

Here is a simple version of the algorithm:

1. Partition the data into training and validation sets.

2. Grow the tree with the training data.

3. Prune it successively, step by step, recording CP (using the training data) at each step.

4. Note the CP that corresponds to the minimum error on the validation data.

5. Repartition the data into training and validation, and repeat the growing, pruning and CP recording process.

6. Do this again and again, and average the CP’s that reflect minimum error for each tree.

7. Go back to the original data, or future data, and grow a tree, stopping at this optimum CP value.

Typically, cross-validation is done such that the partitions (also called “folds”) used for validation are non-overlapping. R automatically does this cross- validation process to select CP and build a default pruned tree, but the user can, instead, specify an alternate value of CP (e.g., if you wanted to see what a deeper tree looked like).

224 CLASSIFICATION AND REGRESSION TREES

In Table 9.4, we see the complexity-parameter table of cross-validation errors for eight trees of increasing depth grown on the Universal Bank data. We can simply choose the tree with the lowest cross-validation error (xerror). In this case, the tree in row 6 has the lowest cross-validation error. Figure 9.12 displays the pruned tree with 15 terminal nodes. This tree was pruned back from the largest tree using the complexity parameter value CP = 0.003436426, which yielded the lowest cross-validation error in Table 9.4.

TABLE 9.4 TABLE OF COMPLEXITY PARAMETER (CP) VALUES AND ASSOCIATED TREE ERRORS

code for tabulating tree error as a function of the complexity parameter (CP)

# argument xval refers to the number of folds to use in rpart's built-in # cross-validation procedure # argument cp sets the smallest value for the complexity parameter. cv.ct <- rpart(Personal.Loan ~ ., data = train.df, method = "class",

cp = 0.00001, minsplit = 5, xval = 5) # use printcp() to print the table. printcp(cv.ct)

Output

CP nsplit rel error xerror xstd 1 0.3350515 0 1.000000 1.00000 0.055705 2 0.1340206 2 0.329897 0.37457 0.035220 3 0.0154639 3 0.195876 0.19931 0.025917 4 0.0068729 7 0.134021 0.17182 0.024096 5 0.0051546 12 0.099656 0.17182 0.024096 6 0.0034364 14 0.089347 0.16838 0.023858 7 0.0022910 19 0.072165 0.17182 0.024096 8 0.0000100 25 0.058419 0.17182 0.024096

Best-Pruned Tree

A further enhancement, in the interest of model parsimony, is to incorporate the sampling error which might cause this minimum to vary if we had a differ- ent sample. The enhancement uses the estimated standard error of the cross- validation error (xstd) to prune the tree even further; we can add one standard error to the minimum xerror. This is sometimes called the Best-Pruned Tree. For example, xstd for the tree in row 6 is 0.023858. We can choose a smaller tree by going up to the row with a cross-validation error that is larger, but still within one standard error—that is, xerror plus xstd (0.16838 + 0.023858 = 0.192238). Here, the tree in row 4 is a smaller tree with the lowest xerror in this range. The best-pruned tree for the loan acceptance example is shown in Figure 9.13. In this case it coincides with the default tree (Figure 9.9).

AVOIDING OVERFITTING 225

In co

m e

< 1

14

C C

A vg

< 3

In co

m e

< 1

06 F am

ily <

3 .5 C C

A vg

> =

1 .1

C D

.A cc

ou nt

< 0

.5

In co

m e

< 9

2

A ge

> =

3 0

Fa m

ily <

2 .5

E du

ca tio

n <

1 .5

In co

m e

< 1

04

A ge

< 4

6

C C

A vg

> =

4

M or

tg ag

e <

1 96

E du

ca tio

n <

1 .5

Fa m

ily <

2 .5

In co

m e

< 1

16

C C

A vg

< 2

.2

O nl

in e

> =

0 .5

27 09

2 91

23 60

5 0

22 33

9

21 83

0

50 9

39 4

11 5

11 2

0 3

12 7

4 1

12 3

3 1

89 1

0

88 7

1 3

34 2

1

32 7

26 1

6 6

6 2

5 0

1 2

0 4

2 1

4

4 1

0

4 4

4 1

0 3

0 6

34 9

2 41

34 2

3 9

34 2

0 0

3 9

7 2

02

7 1

0

6 2

5 0

1 2

1 8

0 1

92

0

0

0

0

0

0

0

0 1

0

0

0

0 1

0

0

0

0

0

0 1

11

1

0

0 1

1

0

0

0 1

1

1

0

0 1

1

1

ye s

no

F IG

U R E

9 .1

2 PR

U N ED

CL AS

SI FI CA

TI ON

TR EE

FO R TH

E LO

AN AC

CE PT AN

CE DA

TA U SI N G CP

TH AT

YI EL DE

D LO

W ES T XE

RR OR

IN TA BL

E 9. 4

code for pruning the tree

# prune by lower cp pruned.ct <- prune(cv.ct,

cp = cv.ct$cptable[which.min(cv.ct$cptable[,"xerror"]),"CP"]) length(pruned.ct$frame$var[pruned.ct$frame$var == "<leaf>"]) prp(pruned.ct, type = 1, extra = 1, split.font = 1, varlen = -10)

226 CLASSIFICATION AND REGRESSION TREES

Income < 114

CCAvg < 3

CD.Account < 0.5

Income < 92

Family < 2.5

Education < 1.5

Family < 2.5

2709 291

2360 50

2233 9

127 41

123 31

89 10

34 21

32 7 2 14

4 10

349 241

342 39

342 0 0 39

7 202

0

0

0

0

0

0

0

0 1

1

0

0

0 1

1

yes no

FIGURE 9.13 BEST-PRUNED TREE OBTAINED BY FITTING A FULL TREE TO THE TRAINING DATA, PRUNING IT USING THE CROSS-VALIDATION DATA, AND CHOOSING THE SMALLEST TREE WITHIN ONE STANDARD ERROR OF THE MINIMUM XERROR TREE

9.5 Classification Rules from Trees

As described in Section 9.1, classification trees provide easily understandable classification rules (if the trees are not too large). Each terminal node is equivalent to a classification rule. Returning to the example, the right-most terminal node in the best-pruned tree (Figure 9.13) gives us the rule

IF (Income ≥ 114) AND (Education ≥ 1.5) THEN Class = 1.

However, in many cases, the number of rules can be reduced by removing redun- dancies. For example, consider the rule from the second-from-bottom-left-most terminal node in Figure 9.13:

IF (Income < 114) AND (CCAvg ≥ 3) AND (CD.Account < 0.5) AND (Income < 92) THEN Class = 0

This rule can be simplified to

IF (Income < 92) AND (CCAvg ≥ 3) AND (CD.Account < 0.5) THEN Class = 0

This transparency in the process and understandability of the algorithm that leads to classifying a record as belonging to a certain class is very advantageous in settings where the final classification is not the only thing of interest. Berry

CLASSIFICATION TREES FOR MORE THAN TWO CLASSES 227

and Linoff (2000) give the example of health insurance underwriting, where the insurer is required to show that coverage denial is not based on discrimination. By showing rules that led to denial (e.g., income < $20K AND low credit history), the company can avoid law suits. Compared to the output of other classifiers, such as discriminant functions, tree-based classification rules are easily explained to managers and operating staff. Their logic is certainly far more transparent than that of weights in neural networks!

9.6 Classification Trees for More Than Two Classes

Classification trees can be used with an outcome variable that has more than two classes. In terms of measuring impurity, the two measures presented earlier (the Gini impurity index and the entropy measure) were defined for m classes and hence can be used for any number of classes. The tree itself would have the same structure, except that its terminal nodes would take one of the m–class labels.

9.7 Regression Trees

The tree method can also be used for a numerical outcome variable. Regression trees for prediction operate in much the same fashion as classification trees. The outcome variable (Y ) is a numerical variable in this case, but both the principle and the procedure are the same: Many splits are attempted, and for each, we measure “impurity” in each branch of the resulting tree. The tree procedure then selects the split that minimizes the sum of such measures. To illustrate a regression tree, consider the example of predicting prices of Toyota Corolla automobiles (from Chapter 6). The dataset includes information on 1000 sold Toyota Corolla cars (We use the first 1000 cars from the dataset ToyotoCorolla.csv. The goal is to find a predictive model of price as a function of 10 predictors (including mileage, horsepower, number of doors, etc.). A regression tree for these data was built using a training set of 600 records. The best-pruned tree is shown in Figure 9.14.

We see that from the 12 input variables (including dummies), only three predictors show up as useful for predicting price: the age of the car, its weight, and horsepower.

Three details differ between regression trees and classification trees: pre- diction, impurity measures, and evaluating performance. We describe these next.

228 CLASSIFICATION AND REGRESSION TREES

Age_08_0 >= 32

Age_08_0 >= 54

Age_08_0 >= 44

Weight < 1258

HP < 98

11756

10386

9358

11696

10952 12332

18443

17811

16417 18882

24262

yes no

FIGURE 9.14 BEST-PRUNED REGRESSION TREE FOR TOYOTA COROLLA PRICES

Prediction

Predicting the outcome value for a record is performed in a fashion similar to the classification case: The predictor information is used for “dropping” the record down the tree until reaching a terminal node. For instance, to predict the price of a Toyota Corolla with Age = 60, Horse_Power = 100, and Weight = 1200, we drop it down the tree and reach the node that has the value $9358. This is the price prediction for this car according to the tree. In classification trees, the value of the terminal node (which is one of the categories) is determined by the “voting” of the training records that were in that terminal node. In regression trees, the value of the terminal node is determined by the average outcome value of the training records that were in that terminal node. In the example above, the value $9358 is the average of the 279 cars in the training set that fall in the category of Age ≥ 54.

Measuring Impurity

We described two types of impurity measures for nodes in classification trees: the Gini index and the entropy-based measure. In both cases, the index is a function of the ratio between the categories of the records in that node. In regression trees, a typical impurity measure is the sum of the squared deviations from the mean of the terminal node. This is equivalent to the sum of the squared errors, because the mean of the terminal node is exactly the prediction. In the example above, the impurity of the node with the value $9358 is computed by subtracting 9358 from the price of each of the 279 cars in the training set that

IMPROVING PREDICTION: RANDOM FORESTS AND BOOSTED TREES 229

fell in that terminal node, then squaring these deviations and summing them up. The lowest impurity possible is zero, when all values in the node are equal.

Evaluating Performance

As stated above, predictions are obtained by averaging the outcome values in the nodes. We therefore have the usual definition of predictions and errors. The predictive performance of regression trees can be measured in the same way that other predictive methods are evaluated (e.g., linear regression), using summary measures such as RMSE.

9.8 Improving Prediction: Random Forests and Boosted Trees

Notwithstanding the transparency advantages of a single tree as described above, in a pure prediction application, where visualizing a set of rules does not matter, better performance is provided by several extensions to trees that combine results from multiple trees. These are examples of ensembles (see Chapter 13). One popular multitree approach is random forests, introduced by Breiman and Cutler.1

Random forests are a special case of bagging, a method for improving predictive power by combining multiple classifiers or prediction algorithms. See Chapter 13 for further details on bagging.

Random Forests

The basic idea in random forests is to:

1. Draw multiple random samples, with replacement, from the data (this sampling approach is called the bootstrap).

2. Using a random subset of predictors at each stage, fit a classification (or regression) tree to each sample (and thus obtain a “forest”).

3. Combine the predictions/classifications from the individual trees to obtain improved predictions. Use voting for classification and averaging for prediction.

The code and output in Figure 9.15 illustrates applying a random forest in R to the personal loan example. The accuracy of the random forest is slightly higher than the single default tree that we fit earlier (compare to the validation performance in Table 9.3).

Unlike a single tree, results from a random forest cannot be displayed in a tree-like diagram, thereby losing the interpretability that a single tree provides.

1For further details on random forests, see www.stat.berkeley.edu/users/breiman/RandomForests/cc_home.htm.

230 CLASSIFICATION AND REGRESSION TREES

code for running a random forest, plotting variable importance plot, and computing accuracy

library(randomForest) ## random forest rf <- randomForest(as.factor(Personal.Loan) ~ ., data = train.df, ntree = 500,

mtry = 4, nodesize = 5, importance = TRUE)

## variable importance plot varImpPlot(rf, type = 1)

## confusion matrix rf.pred <- predict(rf, valid.df) confusionMatrix(rf.pred, valid.df$Personal.Loan)

Partial Output

> confusionMatrix(rf.pred, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1801 19 1 10 170

Accuracy : 0.986

Securities.Account Online CreditCard Mortgage Age CD.Account Experience CCAvg Family Education Income

0 20 40 60 80 100 120 140

rf

MeanDecreaseAccuracy

FIGURE 9.15 VARIABLE IMPORTANCE PLOT FROM RANDOM FOREST (PERSONAL LOAN EXAMPLE)

However, random forests can produce “variable importance” scores, which mea- sure the relative contribution of the different predictors. The importance score for a particular predictor is computed by summing up the decrease in the Gini index for that predictor over all the trees in the forest. Figure 9.15 shows the vari- able importance plots generated from the random forest model for the personal loan example. We see that Income and Education have the highest scores, with

IMPROVING PREDICTION: RANDOM FORESTS AND BOOSTED TREES 231

Family being third. Importance scores for the other predictors are considerably lower.

Boosted Trees

The second type of multitree improvement is boosted trees. Here a sequence of trees is fitted, so that each tree concentrates on misclassified records from the previous tree.

1. Fit a single tree.

2. Draw a sample that gives higher selection probabilities to misclassified records.

3. Fit a tree to the new sample.

4. Repeat Steps 2 and 3 multiple times.

5. Use weighted voting to classify records, with heavier weight for later trees.

Table 9.5 shows the result of running a boosted tree on the loan acceptance example that we saw earlier. We can see that compared to the performance of the single pruned tree (Table 9.3), the boosted tree has better performance on the validation data in terms of overall accuracy and especially in terms of cor- rect classification of 1’s—the rare class of special interest. Where does boosting’s special talent for finding 1’s come from? When one class is dominant (0’s con- stitute over 90% of the data here), basic classifiers are tempted to classify cases as belonging to the dominant class, and the 1’s in this case constitute most of

TABLE 9.5 BOOSTED TREE: CONFUSION MATRIX FOR THE VALIDATION SET (LOAN DATA)

code for running boosted trees

library(adabag) library(rpart) library(caret)

boost <- boosting(Personal.Loan ~ ., data = train.df) pred <- predict(boost, valid.df) confusionMatrix(pred$class, valid.df$Personal.Loan)

Output

Confusion Matrix and Statistics

Reference Prediction 0 1

0 1805 15 1 6 174

Accuracy : 0.9895

232 CLASSIFICATION AND REGRESSION TREES

the misclassifications with the single best-pruned tree. The boosting algorithm concentrates on the misclassifications (which are mostly 1’s), so it is naturally going to do well in reducing the misclassification of 1’s (from 18 in the single tree to 15 in the boosted tree, in the validation set).

9.9 Advantages and Weaknesses of a Tree

Tree methods are good off-the-shelf classifiers and predictors. They are also useful for variable selection, with the most important predictors usually showing up at the top of the tree. Trees require relatively little effort from users in the following senses: First, there is no need for transformation of variables (any monotone transformation of the variables will give the same trees). Second, variable subset selection is automatic since it is part of the split selection. In the loan example, note that the best-pruned tree has automatically selected just three variables (Income, Education, and Family) out of the set of 14 variables available.

Trees are also intrinsically robust to outliers, since the choice of a split depends on the ordering of values and not on the absolute magnitudes of these values. However, they are sensitive to changes in the data, and even a slight change can cause very different splits!

Unlike models that assume a particular relationship between the outcome and predictors (e.g., a linear relationship such as in linear regression and linear discriminant analysis), classification and regression trees are nonlinear and non- parametric. This allows for a wide range of relationships between the predictors and the outcome variable. However, this can also be a weakness: Since the splits are done on one predictor at a time, rather than on combinations of pre- dictors, the tree is likely to miss relationships between predictors, in particular linear structures like those in linear or logistic regression models. Classifica- tion trees are useful classifiers in cases where horizontal and vertical splitting of the predictor space adequately divides the classes. But consider, for instance, a dataset with two predictors and two classes, where separation between the two classes is most obviously achieved by using a diagonal line (as shown in Figure 9.16). In such cases, a classification tree is expected to have lower performance than methods such as discriminant analysis. One way to improve performance is to create new predictors that are derived from existing predictors, which can capture hypothesized relationships between predictors (similar to interactions in regression models). Random forests are another solution in such situations.

Another performance issue with classification trees is that they require a large dataset in order to construct a good classifier. From a computational aspect, trees can be relatively expensive to grow, because of the multiple sorting involved in computing all possible splits on every variable. Pruning the data using the validation sets adds further computation time.

ADVANTAGES AND WEAKNESSES OF A TREE 233

FIGURE 9.16 SCATTER PLOT DESCRIBING A TWO-PREDICTOR CASE WITH TWO CLASSES. THE BEST SEPARATION IS ACHIEVED WITH A DIAGONAL LINE, WHICH CLASSIFICATION TREES CANNOT DO

Although trees are useful for variable selection, one challenge is that they “favor” predictors with many potential split points. This includes categorical predictors with many categories and numerical predictors with many different values. Such predictors have a higher chance of appearing in a tree. One simplis- tic solution is to combine multiple categories into a smaller set and bin numerical predictors with many values. Alternatively, some special algorithms avoid this problem by using a different splitting criterion [e.g., conditional inference trees in the R package party—see Hothorn et al. (2006)—and QUEST classification trees—see Loh and Shih (1997) and www.math.ccu.edu.tw/~yshih/quest.html].

An appealing feature of trees is that they handle missing data without having to impute values or delete records with missing values. Finally, a very important practical advantage of trees is the transparent rules that they generate. Such transparency is often useful in managerial applications, though this advantage is lost in the ensemble versions of trees (random forests, boosted trees).

234 CLASSIFICATION AND REGRESSION TREES

PROBLEMS

9.1 Competitive Auctions on eBay.com. The file eBayAuctions.csv contains informa- tion on 1972 auctions that transacted on eBay.com during May–June 2004. The goal is to use these data to build a model that will classify auctions as competitive or non- competitive. A competitive auction is defined as an auction with at least two bids placed on the item auctioned. The data include variables that describe the item (auction cat- egory), the seller (his/her eBay rating), and the auction terms that the seller selected (auction duration, opening price, currency, day-of-week of auction close). In addi- tion, we have the price at which the auction closed. The task is to predict whether or not the auction will be competitive. Data Preprocessing. Convert variable Duration into a categorical variable. Split the data into training (60%) and validation (40%) datasets.

a. Fit a classification tree using all predictors, using the best-pruned tree. To avoid overfitting, set the minimum number of records in a terminal node to 50 (in R: minbucket = 50). Also, set the maximum number of levels to be displayed at seven (in R: maxdepth = 7).Write down the results in terms of rules. (Note: If you had to slightly reduce the number of predictors due to software limitations, or for clarity of presentation, which would be a good variable to choose?)

b. Is this model practical for predicting the outcome of a new auction?

c. Describe the interesting and uninteresting information that these rules provide.

d. Fit another classification tree (using the best-pruned tree, with a minimum number of records per terminal node = 50 and maximum allowed number of displayed levels = 7), this time only with predictors that can be used for predicting the outcome of a new auction. Describe the resulting tree in terms of rules. Make sure to report the smallest set of rules required for classification.

e. Plot the resulting tree on a scatter plot: Use the two axes for the two best (quan- titative) predictors. Each auction will appear as a point, with coordinates corre- sponding to its values on those two predictors. Use different colors or symbols to separate competitive and noncompetitive auctions. Draw lines (you can sketch these by hand or use R) at the values that create splits. Does this splitting seem reasonable with respect to the meaning of the two predictors? Does it seem to do a good job of separating the two classes?

f. Examine the lift chart and the confusion matrix for the tree. What can you say about the predictive performance of this model?

g. Based on this last tree, what can you conclude from these data about the chances of an auction obtaining at least two bids and its relationship to the auction settings set by the seller (duration, opening price, ending day, currency)? What would you recommend for a seller as the strategy that will most likely lead to a competitive auction?

9.2 Predicting Delayed Flights. The file FlightDelays.csv contains information on all commercial flights departing the Washington, DC area and arriving at New York during January 2004. For each flight, there is information on the departure and arrival airports, the distance of the route, the scheduled time and date of the flight, and so on. The variable that we are trying to predict is whether or not a flight is delayed. A delay is defined as an arrival that is at least 15 minutes later than scheduled.

PROBLEMS 235

Data Preprocessing. Transform variable day of week (DAY_WEEK) info a cate- gorical variable. Bin the scheduled departure time into eight bins (in R use function cut()). Use these and all other columns as predictors (excluding DAY_OF_MONTH). Partition the data into training and validation sets.

a. Fit a classification tree to the flight delay variable using all the relevant predictors. Do not include DEP_TIME (actual departure time) in the model because it is unknown at the time of prediction (unless we are generating our predictions of delays after the plane takes off, which is unlikely). Use a pruned tree with maximum of 8 levels, setting cp = 0.001. Express the resulting tree as a set of rules.

b. If you needed to fly between DCA and EWR on a Monday at 7:00 AM, would you be able to use this tree? What other information would you need? Is it available in practice? What information is redundant?

c. Fit the same tree as in (a), this time excluding the Weather predictor. Display both the pruned and unpruned tree. You will find that the pruned tree contains a single terminal node.

i. How is the pruned tree used for classification? (What is the rule for classifying?)

ii. To what is this rule equivalent?

iii. Examine the unpruned tree. What are the top three predictors according to this tree?

iv. Why, technically, does the pruned tree result in a single node?

v. What is the disadvantage of using the top levels of the unpruned tree as opposed to the pruned tree?

vi. Compare this general result to that from logistic regression in the example in Chapter 10. What are possible reasons for the classification tree’s failure to find a good predictive model?

9.3 Predicting Prices of Used Cars (Regression Trees). The file ToyotaCorolla.csv contains the data on used cars (Toyota Corolla) on sale during late summer of 2004 in the Netherlands. It has 1436 records containing details on 38 attributes, including Price, Age, Kilometers, HP, and other specifications. The goal is to predict the price of a used Toyota Corolla based on its specifications. (The example in Section 9.7 is a subset of this dataset). Data Preprocessing. Split the data into training (60%), and validation (40%) datasets.

a. Run a regression tree (RT) with outcome variable Price and predictors Age_08_04, KM, Fuel_Type, HP, Automatic, Doors, Quarterly_Tax, Mfg_Guarantee, Guarantee_Period, Airco, Automatic_Airco, CD_Player, Powered_Windows, Sport_Model, and Tow_Bar. Keep the minimum number of records in a terminal node to 1, maximum number of tree levels to 100, and cp = 0.001, to make the run least restrictive.

i. Which appear to be the three or four most important car specifications for predicting the car’s price?

ii. Compare the prediction errors of the training and validation sets by examining their RMS error and by plotting the two boxplots. What is happening with the training set predictions? How does the predictive performance of the validation set compare to the training set? Why does this occur?

iii. How can we achieve predictions for the training set that are not equal to the actual prices?

236 CLASSIFICATION AND REGRESSION TREES

iv. Prune the full tree using the cross-validation error. Compared to the full tree, what is the predictive performance for the validation set?

b. Let us see the effect of turning the price variable into a categorical variable. First, create a new variable that categorizes price into 20 bins. Now repartition the data keeping Binned_Price instead of Price. Run a classification tree with the same set of input variables as in the RT, and with Binned_Price as the output variable. Keep the minimum number of records in a terminal node to 1.

i. Compare the tree generated by the CT with the one generated by the RT. Are they different? (Look at structure, the top predictors, size of tree, etc.) Why?

ii. Predict the price, using the RT and the CT, of a used Toyota Corolla with the specifications listed in Table 9.6.

TABLE 9.6 SPECIFICATIONS FOR A PARTICULAR TOYOTA COROLLA

Variable Value

Age_-08_-04 77 KM 117,000 Fuel_Type Petrol HP 110 Automatic No Doors 5 Quarterly_Tax 100 Mfg_Guarantee No Guarantee_Period 3 Airco Yes Automatic_Airco No CD_Player No Powered_Windows No Sport_Model No Tow_Bar Yes

iii. Compare the predictions in terms of the predictors that were used, the mag- nitude of the difference between the two predictions, and the advantages and disadvantages of the two methods.

CHAPTER 10

Logistic Regression

In this chapter, we describe the highly popular and powerful classification method called logistic regression. Like linear regression, it relies on a specific model relating the predictors with the outcome. The user must specify the pre- dictors to include as well as their form (e.g., including any interaction terms). This means that even small datasets can be used for building logistic regression classifiers, and that once the model is estimated, it is computationally fast and cheap to classify even large samples of new records. We describe the logistic regression model formulation and its estimation from data. We also explain the concepts of “logit,” “odds,” and “probability” of an event that arise in the logistic model context and the relations among the three. We discuss variable importance using coefficient and statistical significance and also mention variable selection algorithms for dimension reduction. Our presentation is strictly from a data mining perspective, where classification is the goal and performance is evaluated on a separate validation set. However, because logistic regression is also heav- ily used in statistical analyses for purposes of inference, we give a brief review of key concepts related to coefficient interpretation, goodness-of-fit evaluation, inference, and multiclass models in the Appendix at the end of this chapter.

10.1 Introduction

Logistic regression extends the ideas of linear regression to the situation where the outcome variable, Y , is categorical. We can think of a categorical variable as dividing the records into classes. For example, if Y denotes a recommendation on holding/selling/buying a stock, we have a categorical variable with three categories. We can think of each of the stocks in the dataset (the records) as belonging to one of three classes: the hold class, the sell class, and the buy class.

237

238 LOGISTIC REGRESSION

Logistic regression can be used for classifying a new record, where its class is unknown, into one of the classes, based on the values of its predictor variables (called classification). It can also be used in data where the class is known, to find factors distinguishing between records in different classes in terms of their predictor variables, or “predictor profile” (called profiling). Logistic regression is used in applications such as

1. Classifying customers as returning or nonreturning (classification)

2. Finding factors that differentiate between male and female top executives (profiling)

3. Predicting the approval or disapproval of a loan based on information such as credit scores (classification)

The logistic regression model is used in a variety of fields: whenever a structured model is needed to explain or predict categorical (in particular, binary) outcomes. One such application is in describing choice behavior in econometrics.

In this chapter, we focus on the use of logistic regression for classification. We deal only with a binary outcome variable having two possible classes. In the Appendix, we show how the results can be extended to the case where Y assumes more than two possible classes. Popular examples of binary outcomes are success/failure, yes/no, buy/don’t buy, default/don’t default, and survive/die. For convenience, we often code the values of the binary outcome variable Y as 0 and 1.

Note that in some cases we may choose to convert a continuous outcome variable or an outcome variables with multiple classes into a binary outcome vari- able for purposes of simplification, reflecting the fact that decision-making may be binary (approve the loan/don’t approve, make an offer/don’t make an offer). As with multiple linear regression, the predictor variables X1, X2, . . . , Xk may be categorical variables, continuous variables, or a mixture of these two types. While in multiple linear regression the aim is to predict the value of the contin- uous Y for a new record, in logistic regression the goal is to predict which class a new record will belong to, or simply to classify the record into one of the classes. In the stock example, we would want to classify a new stock into one of the three recommendation classes: sell, hold, or buy. Or, we might want to compute for a new record its propensity (= the probability) to belong to each class, and then possibly rank a set of new records from highest to lowest propensity in order to act on those with the highest propensity.

In logistic regression, we take two steps: the first step yields estimates of the propensities or probabilities of belonging to each class. In the binary case, we get an estimate of p = P(Y = 1), the probability of belonging to class 1 (which also tells us the probability of belonging to class 0). In the next step, we use a cutoff

THE LOGISTIC REGRESSION MODEL 239

value on these probabilities in order to classify each case into one of the classes. For example, in a binary case, a cutoff of 0.5 means that cases with an estimated probability of P(Y = 1) ≥ 0.5 are classified as belonging to class 1, whereas cases with P(Y = 1) < 0.5 are classified as belonging to class 0. This cutoff does not need to be set at 0.5. When the event in question is a low probability but notable or important event (say, 1 = fraudulent transaction), a lower cutoff may be used to classify more cases as belonging to class 1.

10.2 The Logistic Regression Model

The idea behind logistic regression is straightforward: Instead of using Y directly as the outcome variable, we use a function of it, which is called the logit. The logit, it turns out, can be modeled as a linear function of the predictors. Once the logit has been predicted, it can be mapped back to a probability.

To understand the logit, we take several intermediate steps: First, we look at p = P(Y = 1), the probability of belonging to class 1 (as opposed to class 0). In contrast to the binary variable Y , which only takes the values 0 and 1, p can take any value in the interval [0, 1]. However, if we express p as a linear function of the q predictors1 in the form

p = β0 + β1x1 + β2x2 + · · ·+ βqxq, (10.1)

it is not guaranteed that the right-hand side will lead to values within the interval [0, 1]. The solution is to use a nonlinear function of the predictors in the form

p = 1

1 + e−(β0+β1x1+β2x2+···+βqxq) . (10.2)

This is called the logistic response function. For any values x1, . . . , xq, the right- hand side will always lead to values in the interval [0, 1]. Next, we look at a different measure of belonging to a certain class, known as odds. The odds of belonging to class 1 are defined as the ratio of the probability of belonging to class 1 to the probability of belonging to class 0:

Odds(Y = 1) = p

1− p . (10.3)

This metric is very popular in horse races, sports, gambling, epidemiology, and other areas. Instead of talking about the probability of winning or contacting a disease, people talk about the odds of winning or contacting a disease. How are these two different? If, for example, the probability of winning is 0.5, the odds

1Unlike elsewhere in the book, where p denotes the number of predictors, in this chapter we use q, to avoid confusion with the probability p.

240 LOGISTIC REGRESSION

of winning are 0.5/0.5 = 1. We can also perform the reverse calculation: Given the odds of an event, we can compute its probability by manipulating equation (10.3):

p = odds

1 + odds . (10.4)

Substituting (10.2) into (10.4), we can write the relationship between the odds and the predictors as

Odds(Y = 1) = eβ0+β1x1+β2x2+···+βqxq . (10.5)

This last equation describes a multiplicative (proportional) relationship between the predictors and the odds. Such a relationship is interpretable in terms of percentages, for example, a unit increase in predictor Xj is associated with an average increase of βj×100% in the odds (holding all other predictors constant).

Now, if we take a natural logarithm2 on both sides, we get the standard formulation of a logistic model:

log(odds) = β0 + β1x1 + β2x2 + · · ·+ βqxq. (10.6)

The log(odds), called the logit, takes values from −∞ (very low odds) to ∞ (very high odds).3 A logit of 0 corresponds to even odds of 1 (probability = 0.5). Thus, our final formulation of the relation between the outcome and the predictors uses the logit as the outcome variable and models it as a linear function of the q predictors.

To see the relationship between the probability, odds, and logit of belonging to class 1, look at Figure 10.1, which shows the odds (top) and logit (bottom) as a function of p. Notice that the odds can take any non-negative value, and that the logit can take any real value.

10.3 Example: Acceptance of Personal Loan

Recall the example described in Chapter 9 of acceptance of a personal loan by Universal Bank. The bank’s dataset includes data on 5000 customers. The data include the customer’s response to the last personal loan campaign (Personal Loan), as well as customer demographic information (Age, Income, etc.) and the customer’s relationship with the bank (mortgage, securities account, etc.). See Table 10.1. Among these 5000 customers, only 480 (= 9.6%) accepted the personal loan offered to them in a previous campaign. The goal is to build a model that identifies customers who are most likely to accept the loan offer in future mailings.

2The natural logarithm function is typically denoted ln() or log(). In this book, we use log(). 3We use the terms odds and odds(Y = 1) interchangeably.

EXAMPLE: ACCEPTANCE OF PERSONAL LOAN 241

FIGURE 10.1 (a) ODDS AND (b) LOGIT AS A FUNCTION OF p

Model with a Single Predictor

Consider first a simple logistic regression model with just one predictor. This is conceptually analogous to the simple linear regression model in which we fit a straight line to relate the outcome, Y , to a single predictor, X .

Let us construct a simple logistic regression model for classification of cus- tomers using the single predictor Income. The equation relating the outcome

TABLE 10.1 DESCRIPTION OF PREDICTORS FOR ACCEPTANCE OF PERSONAL LOAN EXAMPLE

Age Customer’s age in completed years Experience Number of years of professional experience Income Annual income of the customer ($000s) Family Size Family size of the customer CCAvg Average spending on credit cards per month ($000s) Education Education Level. 1: Undergrad; 2: Graduate; 3: Advanced/Professional Mortgage Value of house mortgage if any ($000s) Securities Account Coded as 1 if customer has securities account with bank CD Account Coded as 1 if customer has certificate of deposit (CD) account with bank Online Banking Coded as 1 if customer uses Internet banking facilities Credit Card Coded as 1 if customer uses credit card issued by Universal Bank

242 LOGISTIC REGRESSION

variable to the predictor in terms of probabilities is

P(Personal Loan = Yes | Income = x) = 1 1 + e−(β0+β1x)

,

or equivalently, in terms of odds,

Odds(Personal Loan = Yes | Income = x) = eβ0+β1x. (10.7)

The estimated coefficients for the model are b0 = −6.16715 and b1 = 0.03757. So the fitted model is

P(Personal Loan = Yes | Income = x) = 1 1 + e6.16715−0.03757x

. (10.8)

Although logistic regression can be used for prediction in the sense that we predict the probability of a categorical outcome, it is most often used for classification. To see the difference between the two, consider predicting the probability of a customer accepting the loan offer as opposed to classifying the customer as an acceptor/nonacceptor. From Figure 10.2, it can be seen that the loan acceptance probabilities produced by the logistic regression model (the s-shaped curve in Figure 10.2) can yield values between 0 and 1. To end up with classifications into either 1 or 0 (e.g., a customer either accepts the loan offer or not), we need a threshold, or cutoff value (see section on “Propensities and Cutoff for Classification” in Chapter 5). This is true in the case of multiple predictor variables as well.

In the Universal Bank example, in order to classify a new customer as an acceptor/nonacceptor of the loan offer, we use the information on his/her

FIGURE 10.2 PLOT OF DATA POINTS (PERSONAL LOAN AS A FUNCTION OF INCOME) AND THE FITTED LOGISTIC CURVE

EXAMPLE: ACCEPTANCE OF PERSONAL LOAN 243

income by plugging it into the fitted equation in (10.8). This yields an esti- mated probability of accepting the loan offer. We then compare it to the cutoff value. The customer is classified as an acceptor if the probability of his/her accepting the offer is above the cutoff.4

Estimating the Logistic Model from Data: Computing Parameter Estimates

In logistic regression, the relation between Y and the β parameters is nonlinear. For this reason, the β parameters are not estimated using the method of least squares (as in multiple linear regression). Instead, a method called maximum likelihood is used. The idea, in brief, is to find the estimates that maximize the chance of obtaining the data that we have. This requires iterations using a computer program.5

Algorithms to compute the coefficient estimates are less robust than algo- rithms for linear regression. Computed estimates are generally reliable for well- behaved datasets where the number of records with outcome variable values of both 0 and 1 are large; their ratio is “not too close” to either 0 or 1; and when the number of coefficients in the logistic regression model is small relative to the sample size (say, no more than 10%). As with linear regression, collinearity (strong correlation among the predictors) can lead to computational difficulties. Computationally intensive algorithms have been developed recently that circum- vent some of these difficulties. For technical details on the maximum likelihood estimation in logistic regression, see Hosmer and Lemeshow (2000).

To illustrate a typical output from such a procedure, we fit a logistic model to the training set of 3000 Universal Bank customers. The outcome variable is Personal Loan, with Yes defined as the success (this is equivalent to setting the outcome variable to 1 for an acceptor and 0 for a nonacceptor).

Data Preprocessing We start by converting predictor variable Education into a factor variable. In the dataset, it is coded as an integer, taking on val- ues 1, 2, or 3. To turn it into a factor variable, we use the R function factor(). Then when we include this predictor variable in R’s logistic regression, it will

4Here we compared the probability to a cutoff c. If we prefer to look at odds of accepting rather than the probability, an equivalent method is to use the equation in (10.7) and compare the odds to c/(1 − c). If the odds are higher than this number, the customer is classified as an acceptor. If it is lower, we classify the customer as a nonacceptor. 5The method of maximum likelihood ensures good asymptotic (large sample) properties for the esti- mates. Under very general conditions, maximum likelihood estimators are: (1) Consistent—The prob- ability of the estimator differing from the true value approaches zero with increasing sample size, (2) Asymptotically efficient—The variance is the smallest possible among consistent estimators, and (3) Asymptotically normally distributed—This allows us to compute confidence intervals and perform statis- tical tests in a manner analogous to the analysis of multiple linear regression models, provided that the sample size is large.

244 LOGISTIC REGRESSION

automatically create two dummy variables from the factor’s three levels. The logistic regression will only use two of the three levels because using all three would create a multicollinearity issue (see Chapter 6). In total, the logistic regres- sion function in R will include 6 = 2 + 1 + 1 + 1 + 1 dummy variables to describe the five categorical predictors from Table 10.1. Together with the six numerical predictors, we have a total of 12 predictors.

Next, we partition the data randomly into training (60%) and validation (40%) sets. We use the training set to fit a logistic regression model and the validation set to assess the model’s performance.

Estimated Model Table 10.2 presents the output from running a logistic regression using the 12 predictors on the training data.

Ignoring p-values for the coefficients, a model based on all 12 predictors has the estimated logistic equation

Logit(Personal Loan = Yes) = (10.9)

−12.6806− 0.0369 Age + 0.0491 Experience +0.0613 Income + 0.5435 Family + 0.2166 CCAvg

+4.2681 EducationGraduate + 4.4408 EducationAdvanced/Professional

+0.0015 Mortgage − 1.1457 Securities.Account + 4.5856 CD.Account −0.8588 Online − 1.2514 Credit Card

The positive coefficients for the dummy variables EducationGraduate, EducProf, and CD.Account mean that holding a CD account and having graduate or professional education (all marked by 1 in the dummy variables) are associated with higher probabilities of accepting the loan offer. In contrast, having a secu- rities account, using online banking, and owning a Universal Bank credit card are associated with lower acceptance rates. For the continuous predictors, posi- tive coefficients indicate that a higher value on that predictor is associated with a higher probability of accepting the loan offer (e.g., Income: higher-income customers tend more to accept the offer). Similarly, negative coefficients indi- cate that a higher value on that predictor is associated with a lower probability of accepting the loan offer (e.g., Age: older customers are less likely to accept the offer).

Interpreting Results in Terms of Odds (for a Profiling Goal)

Logistic models, when they are appropriate for the data, can give useful informa- tion about the roles played by different predictor variables. For example, suppose we want to know how increasing family income by one unit will affect the prob- ability of loan acceptance. This can be found straightforwardly if we consider not probabilities, but odds.

EXAMPLE: ACCEPTANCE OF PERSONAL LOAN 245

TABLE 10.2 LOGISTIC REGRESSION MODEL FOR LOAN ACCEPTANCE (TRAINING DATA)

code for fitting a logistic regression model

bank.df <- read.csv("UniversalBank.csv") bank.df <- bank.df[ , -c(1, 5)] # Drop ID and zip code columns. # treat Education as categorical (R will create dummy variables) bank.df$Education <- factor(bank.df$Education, levels = c(1, 2, 3),

labels = c("Undergrad", "Graduate", "Advanced/Professional"))

# partition data set.seed(2) train.index <- sample(c(1:dim(bank.df)[1]), dim(bank.df)[1]*0.6) train.df <- bank.df[train.index, ] valid.df <- bank.df[-train.index, ]

# run logistic regression # use glm() (general linear model) with family = "binomial" to fit a logistic # regression. logit.reg <- glm(Personal.Loan ~ ., data = train.df, family = "binomial") options(scipen=999) summary(logit.reg)

Output

Estimate Std. Error z value Pr(>|z|) (Intercept) -12.6805628 2.2903370 -5.537 0.0000000308 *** Age -0.0369346 0.0848937 -0.435 0.66351 Experience 0.0490645 0.0844410 0.581 0.56121 Income 0.0612953 0.0039762 15.416 < 0.0000000000000002 *** Family 0.5434657 0.0994936 5.462 0.0000000470 *** CCAvg 0.2165942 0.0601900 3.599 0.00032 *** EducationGraduate 4.2681068 0.3703378 11.525 < 0.0000000000000002 *** EducationAdvanced/Professional 4.4408154 0.3723360 11.927 < 0.0000000000000002 *** Mortgage 0.0015499 0.0007926 1.955 0.05052 . Securities.Account -1.1457476 0.3955796 -2.896 0.00377 ** CD.Account 4.5855656 0.4777696 9.598 < 0.0000000000000002 *** Online -0.8588074 0.2191217 -3.919 0.0000888005 *** CreditCard -1.2514213 0.2944767 -4.250 0.0000214111 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 1901.71 on 2999 degrees of freedom Residual deviance: 682.19 on 2987 degrees of freedom AIC: 708.19

Number of Fisher Scoring iterations: 8

Recall that the odds are given by

Odds = eβ0+β1x1+β2x2+···+βqxq .

At first, let us return to the single predictor example, where we model a cus- tomer’s acceptance of a personal loan offer as a function of his/her income:

Odds(Personal Loan = Yes | Income) = eβ0+β1Income.

246 LOGISTIC REGRESSION

We can think of the model as a multiplicative model of odds. The odds that a customer with income zero will accept the loan is estimated by e−6.16715+(0.03757)(0) = 0.0021. These are the base case odds. In this exam- ple, it is obviously economically meaningless to talk about a zero income; the value zero and the corresponding base-case odds could be meaningful, how- ever, in the context of other predictors. The odds of accepting the loan with an income of $100K will increase by a multiplicative factor of e(0.039)(100) = 42.8 over the base case, so the odds that such a customer will accept the offer are e−6.16715+(0.03757)(100) = 0.0898.

Suppose that the value of Income, or in general X1, is increased by one unit from x1 to x1 + 1, while the other predictors are held at their current value (x2, . . . , x12). We get the odds ratio

odds(x1 + 1, x2, . . . , x12)

odds(x1, . . . , x12) =

eβ0+β1(x1+1)+β2x2+···+β12x12

eβ0+β1x1+β2x2+···+β12x12 = eβ1 .

This tells us that a single unit increase in X1, holding X2, . . . , X12 constant, is associated with an increase in the odds that a customer accepts the offer by a factor of eβ1 . In other words, eβ1 is the multiplicative factor by which the odds (of belonging to class 1) increase when the value of X1 is increased by 1 unit, holding all other predictors constant. If β1 < 0, an increase in X1 is associated with a decrease in the odds of belonging to class 1, whereas a positive value of β1 is associated with an increase in the odds.

When a predictor is a dummy variable, the interpretation is technically the same but has a different practical meaning. For instance, the coefficient for CD.Account was estimated from the data to be 4.023356. Recall that the refer- ence group is customers not holding a CD account. We interpret this coefficient as follows: e4.023356 = 55.9 are the odds that a customer who has a CD account will accept the offer relative to a customer who does not have a CD account, holding all other variables constant. This means that customers who hold CD accounts at Universal Bank are more likely to accept the offer than customers without a CD account (holding all other variables constant).

The advantage of reporting results in odds as opposed to probabilities is that statements such as those above are true for any value of X1. Unless X1 is a dummy variable, we cannot apply such statements about the effect of increasing X1 by a single unit to probabilities. This is because the result depends on the actual value of X1. So if we increase X1 from, say, 3 to 4, the effect on p, the probability of belonging to class 1, will be different than if we increase X1 from 30 to 31. In short, the change in the probability, p, for a unit increase in a particular predictor variable, while holding all other predictors constant, is not a constant—it depends on the specific values of the predictor variables. We therefore talk about probabilities only in the context of specific records.

EVALUATING CLASSIFICATION PERFORMANCE 247

10.4 Evaluating Classification Performance

The general measures of performance that were described in Chapter 5 are used to assess the logistic model performance. Recall that there are several perfor- mance measures, the most popular being those based on the confusion matrix (accuracy alone or combined with costs) and the lift chart. As in other classifi- cation methods, the goal is to find a model that accurately classifies records to their class, using only the predictor information. A variant of this goal is ranking, or finding a model that does a superior job of identifying the members of a par- ticular class of interest for a set of new records (which might come at some cost to overall accuracy). Since the training data are used for selecting the model, we expect the model to perform quite well for those data, and therefore prefer to test its performance on the validation set. Recall that the data in the validation set were not involved in the model building, and thus we can use them to test the model’s ability to classify data that it has not “seen” before.

To obtain the confusion matrix from a logistic regression analysis, we use the estimated equation to predict the probability of class membership (the propensi- ties) for each record in the validation set, and use the cutoff value to decide on the class assignment of these records. We then compare these classifications to the actual class memberships of these records. In the Universal Bank case, we use the estimated model in equation (10.10) to predict the probability of offer acceptance in a validation set that contains 2000 customers (these data were not used in the modeling step). Technically, this is done by predicting the logit using the estimated model in equation (10.10) and then obtaining the probabilities p through the relation p = elogit/1 + elogit. We then compare these probabilities to our chosen cutoff value in order to classify each of the 2000 validation records as acceptors or nonacceptors.

Table 10.3 shows propensities for the first 5 records in the validation set. Suppose that we use a cutoff of 0.5. We see that the first three customers have a probability of accepting the offer that is lower than the cutoff of 0.5, and there- fore they are classified as nonacceptors (0). And indeed, they were nonacceptors (actual = 0). The fourth and fifth customers’ probability of acceptance is esti- mated by the model to exceed 0.5, and they are therefore classified as acceptors (1). While the fourth customer was indeed an acceptor (actual = 1), our model misclassified the fifth customer as an acceptor, when in fact s/he was a nonac- ceptor (actual = 0).

Another useful tool for assessing model classification performance are the lift (gains) chart and decile-wise lift chart (see Chapter 5). Figure 10.3 illustrates the lift chart obtained for the personal loan offer logistic model using the validation set. The “lift” over the base curve indicates for a given number of cases (read on the x-axis), the additional responders that you can identify by using the model. The same information is portrayed in in Figure 10.3: Taking the 10% of the

248 LOGISTIC REGRESSION

TABLE 10.3 PROPENSITIES FOR THE FIRST FIVE CUSTOMERS IN VALIDATION DATA

code for using logistic regression to generate predicted probabilities

# use predict() with type = "response" to compute predicted probabilities. logit.reg.pred <- predict(logit.reg, valid.df[, -8], type = "response")

# first 5 actual and predicted records data.frame(actual = valid.df$Personal.Loan[1:5], predicted = logit.reg.pred[1:5])

Output

> data.frame(actual = valid.df$Personal.Loan[1:5], + predicted = logit.reg.pred[1:5])

actual predicted 2 0 0.00002707663 6 0 0.00326343313 9 0 0.03966293189 10 1 0.98846040544 11 0 0.59933974797

records that are ranked by the model as “most probable 1’s” yields 7.9 times as many 1’s as would simply selecting 10% of the records at random.

Variable Selection

The next step includes searching for alternative models. One option is to look for simpler models by trying to reduce the number of predictors used. We can also build more complex models that reflect interactions among predictors by creating and including new variables that are derived from the predictors. For example, if we hypothesize that there is an interactive effect between income and family size, we should add an interaction term of the form Income × Family. The choice among the set of alternative models is guided primarily by perfor- mance on the validation data. For models that perform roughly equally well, simpler models are generally preferred over more complex models. Note also that performance on validation data may be overly optimistic when it comes to predicting performance on data that have not been exposed to the model at all. This is because when the validation data are used to select a final model among a set of model, we are selecting based on how well the model performs with those data and therefore may be incorporating some of the random idiosyncrasies of the validation data into the judgment about the best model. The model still may be the best for the validation data among those considered, but it will probably not do as well with the unseen data. Therefore, it is useful to evaluate the chosen model on a new test set to get a sense of how well it will perform on new data. In addition, one must consider practical issues such as costs of collecting variables, error-proneness, and model complexity in the selection of the final model.

EVALUATING CLASSIFICATION PERFORMANCE 249

code for creating lift chart and decile-wise lift chart

library(gains) gain <- gains(valid.df$Personal.Loan, logit.reg.pred, groups=length(logit.reg.pred))

# plot lift chart plot(c(0,gain$cume.pct.of.total*sum(valid.df$Personal.Loan))~c(0,gain$cume.obs),

xlab="# cases", ylab="Cumulative", main="", type="l") lines(c(0,sum(valid.df$Personal.Loan))~c(0, dim(valid.df)[1]), lty=2)

# compute deciles and plot decile-wise chart heights <- gain$mean.resp/mean(valid.df$Personal.Loan) midpoints <- barplot(heights, names.arg = gain$depth, ylim = c(0,9),

xlab = "Percentile", ylab = "Mean Response", main = "Decile-wise lift chart")

# add labels to columns text(midpoints, heights+0.5, labels=round(heights, 1), cex = 0.8)

0 500 1000 1500 2000

0 50

10 0

15 0

# cases

C um

ul at

iv e

10 20 30 40 50 60 70 80 90 100

Decile-wise lift chart

Percentile

M ea

n R

es po

ns e

0 2

4 6

8

7.9

1.1 0.4 0.3 0.1 0 0.1 0.1 0.1 0

FIGURE 10.3 LIFT CHART AND DECILE-WISE LIFT CHART FOR THE VALIDATION DATA FOR UNIVERSAL BANK LOAN OFFER

250 LOGISTIC REGRESSION

As in linear regression, in logistic regression we can use automated variable selection heuristics such as stepwise selection, forward selection, and backward elimination (See Section 6.4 in Chapter 6.) In R, use function step() in the stats package or function stepAIC() in the MASS package) for stepwise, forward, and backward elimination. If the dataset is not too large, we can even try an exhaustive search over all possible models (use R function glmulti() in package glmulti, although it can be slow).

10.5 Example of Complete Analysis: Predicting Delayed Flights

Predicting flight delays can be useful to a variety of organizations: airport author- ities, airlines, aviation authorities. At times, joint task forces have been formed to address the problem. Such an organization, if it were to provide ongoing real-time assistance with flight delays,would benefit from some advance notice about flights likely to be delayed.

In this simplified illustration, we look at six predictors (see Table 10.4). The outcome of interest is whether the flight is delayed or not (delayed means more than 15 minutes late). Our data consist of all flights from the Washington, DC area into the New York City area during January 2004. The percent of delayed flights among these 2201 flights is 19.5%. The data were obtained from the Bureau of Transportation Statistics website (www.transtats.bts.gov).

The goal is to predict accurately whether a new flight, not in this dataset, will be delayed or not. The outcome variable is a variable called Flight Status, coded as delayed or ontime.

Other information available on the website, such as distance and arrival time, is irrelevant because we are looking at a certain route (distance, flight time, etc. should be approximately equal for all flights in the data). A sample of the data for 20 flights is shown in Table 10.5. Figures 10.4 and 10.5 show visualizations of

TABLE 10.4 DESCRIPTION OF PREDICTORS FOR FLIGHT DELAYS EXAMPLE

Day of Week Coded as 1 = Monday, 2 = Tuesday,..., 7 = Sunday Departure Time Broken down into 18 intervals between 6:00 AM and 10:00 PM Origin Three airport codes: DCA (Reagan National), IAD (Dulles),

BWI (Baltimore–Washington Int’l) Destination Three airport codes: JFK (Kennedy), LGA (LaGuardia),

EWR (Newark) Carrier Eight airline codes: CO (Continental), DH (Atlantic Coast),

DL (Delta), MQ (American Eagle), OH (Comair), RU (Continental Express), UA (United), and US (USAirways)

Weather Coded as 1 if there was a weather-related delay

EXAMPLE OF COMPLETE ANALYSIS: PREDICTING DELAYED FLIGHTS 251

TABLE 10.5 SAMPLE OF 20 FLIGHTS

Flight Day of Departure Status Carrier Week Time Destination Origin Weather

ontime DL 2 728 LGA DCA 0 delayed US 3 1600 LGA DCA 0 ontime DH 5 1242 EWR IAD 0 ontime US 2 2057 LGA DCA 0 ontime DH 3 1603 JFK IAD 0 ontime CO 6 1252 EWR DCA 0 ontime RU 6 1728 EWR DCA 0 ontime DL 5 1031 LGA DCA 0 ontime RU 6 1722 EWR IAD 0 delayed US 1 627 LGA DCA 0 delayed DH 2 1756 JFK IAD 0 ontime MQ 6 1529 JFK DCA 0 ontime US 6 1259 LGA DCA 0 ontime DL 2 1329 LGA DCA 0 ontime RU 2 1453 EWR BWI 0 ontime RU 5 1356 EWR DCA 0 delayed DH 7 2244 LGA IAD 0 ontime US 7 1053 LGA DCA 0 ontime US 2 1057 LGA DCA 0 ontime US 4 632 LGA DCA 0

the relationships between flight delays and different predictors or combinations of predictors. From Figure 10.4, we see that Sundays and Mondays saw the largest proportion of delays. Delay rates also seem to differ by carrier, by time of day, as well as by origin and destination airports. For Weather, we see a strong distinction between delays when Weather = 1 (in that case there is always a delay) and Weather = 0. The heatmap in Figure 10.5 reveals some specific combinations with high rates of delays, such as Sunday flights by carrier RU, departing from BWI, or Sunday flights by MQ departing from DCA. We can also see combinations with very low delay rates.

Our main goal is to find a model that can obtain accurate classifications of new flights based on their predictor information. An alternative goal is finding a certain percentage of flights that are most/least likely to get delayed (ranking). And a third different goal is profiling flights: finding out which factors are asso- ciated with a delay (not only in this sample but in the entire population of flights on this route), and for those factors we would like to quantify these effects. A logistic regression model can be used for all these goals, albeit in different ways.

Data Preprocessing

Create a binary outcome variable called isDelay that takes the value 1 if Flight Status = delayed and 0 otherwise. Transform day of week into a categorical

252 LOGISTIC REGRESSION

code for generating bar charts of average delay vs. predictors

# code for generating top-right bar chart # for other plots, replace aggregating variable by setting argument by = in # aggregate(). # in function barplot(), set the x-label (argument xlab =) and y-label # (argument names.arg =) # according to the variable of choice.

barplot(aggregate(delays.df$Flight.Status == "delayed", by = list(delays.df$DAY_WEEK), mean, rm.na = T)[,2], xlab = "Day of Week", ylab = "Average Delay", names.arg = c("Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun"))

Mon Tue Wed Thu Fri Sat Sun

Day of Week

A ve

ra ge

D el

ay

0. 00

0. 10

0. 20

6 8 10 12 14 16 18 20

Departure Time

A ve

ra ge

D el

ay

0. 0

0. 1

0. 2

0. 3

IAD BWI DCA

Origin

A ve

ra ge

D el

ay

0. 00

0. 10

0. 20

LGA EWR JFK

Destination

A ve

ra ge

D el

ay

0. 00

0. 10

0. 20

US CO DH DL MQ OH RU UA

Carrier

A ve

ra ge

D el

ay

0. 00

0. 10

0. 20

0 1

Weather

A ve

ra ge

D el

ay

0. 0

0. 4

0. 8

FIGURE 10.4 PROPORTION OF DELAYED FLIGHTS BY EACH OF THE SIX PREDICTORS. TIME OF DAY IS DIVIDED INTO HOURLY BINS

EXAMPLE OF COMPLETE ANALYSIS: PREDICTING DELAYED FLIGHTS 253

code for generating heatmap for exploring flight delays

library(reshape) library(ggplot2) # create matrix for plot agg <- aggregate(delays.df$Is.Delay,

by = list(delays.df$DAY_WEEK, delays.df$CARRIER, delays.df$ORIGIN), FUN = mean, na.rm = TRUE)

m <- melt(agg) names(m)[1:3] <- c("DAY_WEEK", "CARRIER", "ORIGIN")

# plot with ggplot # use facet_grid() with arguments scales = "free" and space = "free" to skip # missing values. ggplot(m, aes(y = CARRIER, x = DAY_WEEK, fill = value)) + geom_tile() +

facet_grid(ORIGIN ~ ., scales = "free", space = "free") + scale_fill_gradient(low="white", high="black")

OH

RU

CO

DH

DL

MQ

RU

US

DH

RU

UA

B W

I D

C A

IA D

Mon Tue Wed Thu Fri Sat Sun

DAY_WEEK

C A

R R

IE R

0.0

0.1

0.2

0.3

0.4

0.5

value

FIGURE 10.5 PERCENT OF DELAYED FLIGHTS (DARKER = HIGHER %DELAYS) BY DAY OF WEEK, ORIGIN, AND CARRIER

variable, and bin and categorize the departure time into hourly intervals between 6:00 AM and 10:00 PM. Set reference categories for categorical variables: IAD for departure airport, LGA for arrival, USAirways for carrier, and Wednesday for day (see Figure 10.4 for R code). This yields a total of 34 dummies. In addition, we have a single dummy for Weather. While this is a large number of predictors, we start by using all of them, look at performance, and then explore reducing the dimension. We then partition the data into training set (60%) and validation set (40%). We use the training set to fit a model and the validation set to assess the model’s performance.

254 LOGISTIC REGRESSION

Model-Fitting and Estimation

The estimated model with 34 predictors is shown in Table 10.6. Note how negative coefficients in the logit model (the “Coefficient” column) translate into odds coefficients lower than 1, and positive logit coefficients translate into odds coefficients larger than 1.

Model Interpretation

The coefficient for Arrival Airport JFK (DESTJFK) is estimated as −0.19. Recall that the reference group is LGA. We interpret this coefficient as follows: e−0.19 = 0.83 are the odds of a flight arriving at JFK being delayed relative to a flight to LGA being delayed (= the base-case odds), holding all other variables constant. This means that flights to LGA are more likely to be delayed than those to JFK (holding everything else constant). If we consider statistical sig- nificance of the coefficients, we see that in general, the origin and destination airports are not associated with the chance of delays. For carriers, two carriers (CO, MQ) are significantly different from the base carrier (USAirways), with odds of delay ranging between 4 and 5.5 relative to the other airlines. Weather has an enormous coefficient, which is not statistically significant. Flights leaving on Sunday or Monday have, on average, odds of 1.8 of delays relative to other days of the week (the other days seem statistically similar to the reference group Wednesday). Also, odds of delays appear to change over the course of the day, with the most noticeable difference from the reference category (6–7 AM) being 3–4 PM.

Model Performance

How should we measure the performance of models? One possible measure is “percent of flights correctly classified.” Accurate classification can be obtained from the confusion matrix for the validation data. The confusion matrix gives a sense of the classification accuracy and what type of misclassification is more fre- quent. From the confusion matrix and error rates in Figure 10.6, it can be seen that the model more accurately classifies nondelayed flights and is less accurate in classifying flights that were delayed. (Note: The same pattern appears in the confusion matrix for the training data, so it is not surprising to see it emerge for new data.) If there is an asymmetric cost structure so that one type of misclassifi- cation is more costly than the other, the cutoff value can be selected to minimize the cost. Of course, this tweaking should be carried out on the training data and assessed only using the validation data.

In most conceivable situations, the purpose of the model would be to identify those flights most likely to be delayed among a set of flights, so that resources can be directed toward either reducing the delay or mitigating its effects. Air

EXAMPLE OF COMPLETE ANALYSIS: PREDICTING DELAYED FLIGHTS 255

TABLE 10.6 ESTIMATED LOGISTIC REGRESSION MODEL FOR DELAYED FLIGHTS (BASED ON THE TRAINING SET)

code for data preprocessing and running logistic regression

delays.df <- read.csv("FlightDelays.csv")

# transform variables and create bins delays.df$DAY_WEEK <- factor(delays.df$DAY_WEEK, levels = c(1:7),

labels = c("Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun")) delays.df$CRS_DEP_TIME <- factor(round(delays.df$CRS_DEP_TIME/100))

# create reference categories delays.df$ORIGIN <- relevel(delays.df$ORIGIN, ref = "IAD") delays.df$DEST <- relevel(delays.df$DEST, ref = "LGA") delays.df$CARRIER <- relevel(delays.df$CARRIER, ref = "US") delays.df$DAY_WEEK <- relevel(delays.df$DAY_WEEK, ref = "Wed") delays.df$isDelay <- 1 * (delays.df$Flight.Status == "delayed")

# create training and validation sets selected.var <- c(10, 1, 8, 4, 2, 9, 14) train.index <- sample(c(1:dim(delays.df)[1]), dim(delays.df)[1]*0.6) train.df <- delays.df[train.index, selected.var] valid.df <- delays.df[-train.index, selected.var]

# run logistic model, and show coefficients and odds lm.fit <- glm(isDelay ~ ., data = train.df, family = "binomial") data.frame(summary(lm.fit)$coefficients, odds = exp(coef(lm.fit)))

> round(data.frame(summary(lm.fit)$coefficients, odds = exp(coef(lm.fit))), 5) Estimate Std..Error z.value Pr...z.. odds

(Intercept) -2.60619 0.61423 -4.24302 0.00002 7.382000e-02 DAY_WEEKMon 0.52118 0.26915 1.93644 0.05281 1.684020e+00 DAY_WEEKTue 0.14391 0.28607 0.50307 0.61491 1.154780e+00 DAY_WEEKThu -0.05467 0.27159 -0.20128 0.84048 9.468000e-01 DAY_WEEKFri -0.00027 0.27089 -0.00098 0.99922 9.997300e-01 DAY_WEEKSat -0.74215 0.35108 -2.11392 0.03452 4.760900e-01 DAY_WEEKSun 0.66789 0.27670 2.41378 0.01579 1.950120e+00 CRS_DEP_TIME7 -0.04694 0.51691 -0.09081 0.92764 9.541400e-01 CRS_DEP_TIME8 0.29127 0.48747 0.59751 0.55017 1.338130e+00 CRS_DEP_TIME9 -0.36183 0.60420 -0.59886 0.54927 6.964000e-01 CRS_DEP_TIME10 -0.34262 0.59867 -0.57231 0.56711 7.099100e-01 CRS_DEP_TIME11 -0.58081 0.82897 -0.70064 0.48353 5.594500e-01 CRS_DEP_TIME12 0.62069 0.46722 1.32848 0.18402 1.860210e+00 CRS_DEP_TIME13 0.03636 0.51341 0.07082 0.94354 1.037030e+00 CRS_DEP_TIME14 0.25826 0.50724 0.50915 0.61065 1.294680e+00 CRS_DEP_TIME15 1.04599 0.41727 2.50677 0.01218 2.846230e+00 CRS_DEP_TIME16 0.55740 0.45512 1.22474 0.22067 1.746130e+00 CRS_DEP_TIME17 0.67856 0.42316 1.60355 0.10881 1.971040e+00 CRS_DEP_TIME18 0.28052 0.59063 0.47494 0.63483 1.323810e+00 CRS_DEP_TIME19 0.75145 0.50054 1.50126 0.13329 2.120070e+00 CRS_DEP_TIME20 0.92481 0.68216 1.35570 0.17519 2.521380e+00 CRS_DEP_TIME21 0.88665 0.44103 2.01041 0.04439 2.426980e+00 ORIGINBWI 0.40424 0.38936 1.03821 0.29917 1.498170e+00 ORIGINDCA -0.43414 0.36966 -1.17443 0.24022 6.478200e-01 DESTEWR -0.06771 0.33162 -0.20417 0.83822 9.345400e-01 DESTJFK -0.18893 0.25224 -0.74900 0.45386 8.278500e-01 CARRIERCO 1.70665 0.53019 3.21895 0.00129 5.510500e+00 CARRIERDH 0.92781 0.49642 1.86898 0.06163 2.528960e+00 CARRIERDL 0.30436 0.33592 0.90605 0.36491 1.355750e+00 CARRIERMQ 1.38772 0.32314 4.29444 0.00002 4.005710e+00 CARRIEROH -0.40804 0.84431 -0.48327 0.62890 6.649600e-01 CARRIERRU 0.94661 0.49224 1.92309 0.05447 2.576970e+00 CARRIERUA 0.51400 0.83733 0.61385 0.53931 1.671970e+00 Weather 17.85034 502.07524 0.03555 0.97164 5.653314e+07

256 LOGISTIC REGRESSION

code for evaluating performance of all-predictor model

library(gains) pred <- predict(lm.fit, valid.df) gain <- gains(valid.df$isDelay, lm.fit$fitted.values, groups=100)

plot(c(0,gain$cume.pct.of.total*sum(valid.df$isDelay))~ c(0,gain$cume.obs), xlab="# cases", ylab="Cumulative", main="", type="l")

lines(c(0,sum(valid.df$isDelay))~c(0, dim(valid.df)[1]), lty=2)

Output

> confusionMatrix(ifelse(pred > 0.5, 1, 0), valid.df$isDelay) Confusion Matrix and Statistics

Reference Prediction 0 1

0 714 156 1 0 11

Accuracy : 0.8229

0 200 400 600 800

0 50

10 0

15 0

# cases

C um

ul at

iv e

FIGURE 10.6 CONFUSION MATRIX AND LIFT CHART FOR THE FLIGHT DELAY VALIDATION DATA USING ALL PREDICTORS

traffic controllers might work to open up additional air routes or allocate more controllers to a specific area for a short time. Airlines might bring on personnel to rebook passengers and to activate standby flight crews and aircraft. Hotels might allocate space for stranded travellers. In all cases, the resources available are going to be limited and might vary over time and from organization to organization. In this situation, the most useful model would provide an ordering of flights by their probability of delay, letting the model users decide how far down that list to go in taking action. Therefore, model lift is a useful measure of performance— as you move down that list of flights, ordered by their delay probability, how much better does the model do in predicting delay than would a naive model which is simply the average delay rate for all flights? From the lift curve for the

EXAMPLE OF COMPLETE ANALYSIS: PREDICTING DELAYED FLIGHTS 257

validation data (Figure 10.6), we see that our model is superior to the baseline (simple random selection of flights).

Variable Selection

From the data exploration charts (Figures 10.4 and 10.5) and from the coef- ficient table for the flights delay model (Table 10.6), it appears that several of the predictors could be dropped or coded differently. Additionally, we look at the number of flights in different categories to identify categories with very few or no flights—such categories are candidates for removal or merger (see Table 10.7).

First, we find that most carriers depart from a single airport (DCA): For those that depart from all three airports, the delay rates are similar regardless of airport. We therefore drop the departure airport distinction by excluding Origin dummies and find that the model performance and fit is not harmed. We also drop the destination airport for a practical reason: Not all carriers fly to all airports. Our model would then be invalid for prediction in nonexistent combinations of carrier and destination airport. We also try grouping carriers, day of week, and hour of day into fewer categories that are more distinguishable with respect to delays. For example, Sundays and Mondays seem to have a similar rate of delays, which differs from the lower rate on Tuesday–Saturday. We therefore group the days of week into Sunday+Monday and Other, resulting in a single dummy variable.

Table 10.8 displays the estimated smaller model, with its training and valida- tion confusion matrices and error rates. Figure 10.7 presents the lift chart on the validation set. It can be seen that this model competes well with the larger model in terms of classification accuracy and lift, while using much less information.

We therefore conclude with a six-predictor model that requires only knowl- edge of the carrier, the day of week, the hour of the day, and whether it is likely that there will be a delay due to weather. However, this weather vari- able refers to actual weather at flight time, not a forecast, and is not known in advance! If the aim is to predict in advance whether a particular flight will be

TABLE 10.7 NUMBER OF FLIGHTS BY CARRIER AND ORIGIN

BWI DCA IAD Total CO 94 94 DH 27 524 551 DL 388 388 MQ 295 295 OH 30 30 RU 115 162 131 408 UA 31 31 US 404 404

Total 145 1370 686 2201

258 LOGISTIC REGRESSION

TABLE 10.8 LOGISTIC REGRESSION MODEL WITH FEWER PREDICTORS

code for logistic regression with fewer predictors

# fewer predictors delays.df$Weekend <- delays.df$DAY_WEEK %in% c("Sun", "Sat") delays.df$CARRIER_CO_MQ_DH_RU <- delays.df$CARRIER %in% c("CO", "MQ", "DH", "RU") delays.df$MORNING <- delays.df$CRS_DEP_TIME %in% c(6, 7, 8, 9) delays.df$NOON <- delays.df$CRS_DEP_TIME %in% c(10, 11, 12, 13) delays.df$AFTER2P <- delays.df$CRS_DEP_TIME %in% c(14, 15, 16, 17, 18) delays.df$EVENING <- delays.df$CRS_DEP_TIME %in% c(19, 20)

set.seed(1) # Set the seed for the random number generator for reproducing the # partition.

train.index <- sample(c(1:dim(delays.df)[1]), dim(delays.df)[1]*0.6) valid.index <- setdiff(c(1:dim(delays.df)[1]), train.index) train.df <- delays.df[train.index, ] valid.df <- delays.df[valid.index, ]

lm.fit <- glm(isDelay ~ Weekend + Weather + CARRIER_CO_MQ_DH_RU + MORNING + NOON + AFTER2P + EVENING, data = train.df, family = "binomial")

summary(lm.fit)

# evaluate pred <- predict(lm.fit, valid.df) confusionMatrix(ifelse(pred > 0.5, 1, 0), valid.df$isDelay)

Output

> summary(lm.fit)

Coefficients: Estimate Std. Error z value Pr(>|z|)

(Intercept) -1.974836 0.287812 -6.862 6.81e-12 *** WeekendTRUE -0.008063 0.173019 -0.047 0.9628 Weather 18.100254 501.012404 0.036 0.9712 CARRIER_CO_MQ_DH_RUTRUE 1.130216 0.178381 6.336 2.36e-10 *** MORNINGTRUE -0.739863 0.292933 -2.526 0.0115 * NOONTRUE -0.567475 0.298423 -1.902 0.0572 . AFTER2PTRUE -0.106968 0.259789 -0.412 0.6805 EVENINGTRUE 0.181173 0.357225 0.507 0.6120 --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

> confusionMatrix(ifelse(pred > 0.5, 1, 0), valid.df$isDelay) Confusion Matrix and Statistics

Reference Prediction 0 1

0 714 156 1 0 11

Accuracy : 0.8229

APPENDIX: LOGISTIC REGRESSION FOR PROFILING 259

0 200 400 600 800

0 50

10 0

15 0

# cases

C um

ul at

iv e

FIGURE 10.7 LIFT CHART FOR LOGISTIC REGRESSION MODEL WITH FEWER PREDICTORS ON THE VALIDATION SET

delayed, a model without Weather must be used. In contrast, if the goal is profiling delayed vs. ontime flights, we can keep Weather in the model to allow evaluating the impact of the other factors while holding weather constant [that is, (approximately) comparing days with weather delays to days without weather delays].

To conclude, based on the model built from January 2004 data, the highest chance of an ontime flight from DC to New York is on Tuesday–Saturday around noon, on Delta, Comair, United, or USAirways. And clearly, good weather is advantageous!

10.6 Appendix: Logistic Regression for Profiling

The presentation of logistic regression in this chapter has been primarily from a data mining perspective where classification or ranking is the goal, and perfor- mance is evaluated by reviewing results with a validation sample. For reference, some key concepts of a classical statistical perspective are included below.

Appendix A: Why Linear Regression Is Problematic for a Categorical Outcome

Now that you have seen how logistic regression works, we explain why it is considered preferable over linear regression for a binary outcome. Technically, one can apply a multiple linear regression model to this problem, treating the outcome variable Y as continuous. This is called a Linear Probability Model. Of course, Y must be coded numerically (e.g., 1 for customers who accepted the loan offer and 0 for customers who did not accept it). Although software will

260 LOGISTIC REGRESSION

yield an output that at first glance may seem standard (e.g., Table 10.9), a closer look will reveal several anomalies:

1. Using the model to predict Y for each of the records (or classify them) yields predictions that are not necessarily 0 or 1.

2. A look at the histogram or probability plot of the residuals reveals that the assumption that the outcome variable (or residuals) follows a normal distribution is violated. Clearly, if Y takes only the values 0 and 1, it cannot be normally distributed. In fact, a more appropriate distribution for the number of 1’s in the dataset is the binomial distribution with p = P (Y = 1).

3. The assumption that the variance of Y is constant across all classes is violated. Since Y follows a binomial distribution, its variance is np(1− p). This means that the variance will be higher for classes where the probability of adoption, p, is near 0.5 than where it is near 0 or 1.

The first anomaly is the main challenge when the goal is classification, especially if we are interested in propensities (p = P (Y = 1)). The second and third anomalies are relevant to profiling, where we use statistical inference that relies on standard errors.

Below you will find partial output from running a multiple linear regression of Personal Loan (PL, coded as PL = 1 for customers who accepted the loan offer and PL = 0 otherwise) on three of the predictors. The estimated model is

P̂L = −0.2326196 + 0.0030989 Income + 0.0344897 Family + 0.2872284 CD

TABLE 10.9 OUTPUT FOR MULTIPLE LINEAR REGRESSION MODEL OF PERSONAL LOAN ON THREE PREDICTORS

> reg <- lm(Personal.Loan ~ Income + Family + CD.Account, data = bank.df) > summary(reg)

Call: lm(formula = Personal.Loan ~ Income + Family + CD.Account, data = bank.df)

Residuals: Min 1Q Median 3Q Max

-0.81774 -0.12536 -0.02930 0.06407 0.99670

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) -0.2326196 0.0103867 -22.40 <0.0000000000000002 *** Income 0.0030989 0.0000765 40.51 <0.0000000000000002 *** Family 0.0344897 0.0030243 11.40 <0.0000000000000002 *** CD.Account 0.2872284 0.0145981 19.68 <0.0000000000000002 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.2421 on 4996 degrees of freedom Multiple R-squared: 0.325, Adjusted R-squared: 0.3246 F-statistic: 801.9 on 3 and 4996 DF, p-value: < 0.00000000000000022

APPENDIX: LOGISTIC REGRESSION FOR PROFILING 261

To predict whether a new customer will accept the personal loan offer (PL = 1) or not (PL = 0), we input the information on its values for these three pre- dictors. For example, we would predict the loan offer acceptance of a cus- tomer with an annual income of $50K with two family members, who does not hold CD accounts in Universal Bank to be −0.2326196+ (0.0030989)(50)+ (0.0344897)(2) = −0.0086952. Clearly, this is not a valid “loan acceptance” value. Furthermore, the histogram of the residuals (Figure 10.8) reveals that the residuals are probably not normally distributed. Therefore, our estimated model is based on violated assumptions and cannot be used for inference.

Residuals

F re

qu en

cy

−0.5 0.0 0.5 1.0

0 40

0 80

0 12

00

FIGURE 10.8 HISTOGRAM OF RESIDUALS FROM A MULTIPLE LINEAR REGRESSION FOR UNIVERSAL BANK ON THE THREE PREDICTORS. WE SEE THE RESIDUALS DO NOT FOLLOW THE NORMAL DISTRIBUTION

Appendix B: Evaluating Explanatory Power

When the purpose of the analysis is profiling (identifying predictor profiles that distinguish the two classes, or explaining the differences between the classes in terms of predictor values), we are less interested in how well the model classi- fies new data than in how well the model fits the data it was trained on. For example, if we are interested in characterizing the average loan offer acceptor vs. nonacceptor in terms of income, education, and so on, we want to find a model that fits the data best. We therefore mention popular measures used to assess how well the model fits the data. Clearly, we look at the training set in order to evaluate goodness of fit (and in fact, we do not need to partition the data).

Overall Strength-of-Fit As in multiple linear regression, we first evaluate the overall explanatory power of the model before looking at single predictors.

262 LOGISTIC REGRESSION

We ask: Is this set of predictors better than a simple naive model for explaining the difference between classes?6

The deviance D is a statistic that measures overall goodness of fit. It is similar to the concept of sum of squared errors (SSE) in the case of least squares estimation (used in linear regression). We compare the deviance of our model, D (called Residual deviance in R), to the deviance of the naive (Null) model, D0. For example, in Table 10.10, we see D = 682.19 and D0 = 1901.71. If the reduction in deviance is statistically significant (as indicated by a low p-value7), we consider our model to provide a good overall fit, better than a model with no explanatory (X) variables.

TABLE 10.10 MEASURES OF EXPLANATORY POWER FOR UNIVERSAL BANK TRAINING DATA WITH A 12 PREDICTOR MODEL

> summary(logit.reg)

Call: glm(formula = Personal.Loan ~ ., family = "binomial", data = train.df)

Deviance Residuals: Min 1Q Median 3Q Max

-2.0380 -0.1847 -0.0627 -0.0183 3.9810

Coefficients: [OMITTED] --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 1901.71 on 2999 degrees of freedom Residual deviance: 682.19 on 2987 degrees of freedom AIC: 708.19

Number of Fisher Scoring iterations: 8

Finally, the confusion matrix and lift chart for the training data (Figure 10.9) give a sense of how accurately the model classifies the data. If the model fits the data well, we expect it to classify these data accurately into their actual classes.

Impact of Single Predictors As in multiple linear regression, the output from a logistic regression procedure typically yields a coefficient table, where

6In a naive model, no explanatory variables (X ’s) exist and each record is classified as belonging to the majority class. 7The difference between the deviance of a naive model and deviance of the model at hand approxi- mately follows a chi-squared distribution with k degrees of freedom, where k is the number of predic- tors in the model at hand. Therefore, to get the p-value, compute the difference between the deviances (d) and then compute the probability that a chi-squared variable with k degrees of freedom is larger than d.

APPENDIX: LOGISTIC REGRESSION FOR PROFILING 263

> confusionMatrix(ifelse(logit.reg$fitted > 0.5, 1, 0), train.df[, 8]) Confusion Matrix and Statistics

Reference Prediction 0 1

0 2687 92 1 24 197

Accuracy : 0.9613

0 500 1000 1500 2000

0 50

10 0

15 0

# cases

C um

ul at

iv e

FIGURE 10.9 CONFUSION MATRIX AND LIFT CHART FOR UNIVERSAL BANK TRAINING DATA WITH 12 PREDICTORS

for each predictor Xi, we have an estimated coefficient bi and an associated standard error. The associated p-value indicates the statistical significance of the predictor Xi, with very low p-values indicating a statistically significant relation- ship between the predictor and the outcome (given that the other predictors are accounted for), a relationship that is most likely not a result of chance. Three important points to remember are:

1. A statistically significant relationship is not necessarily a practically significant one, in which the predictor has great impact. If the sample is very large, the p-value will be very small simply because the chance uncertainty associated with fitting a model to a particular data set, which can be considerable in a small sample, is negligible in a large sample.

2. Comparing the coefficient magnitudes, or equivalently the odds mag- nitudes, is meaningless unless all predictors have the same scale. Recall that each coefficient is multiplied by the predictor value, so that different predictor scales lead to different coefficient scales.

264 LOGISTIC REGRESSION

3. A statistically significant predictor means that on average, a unit increase in that predictor is associated with a certain effect on the outcome (holding all other predictors constant). It does not, however, indicate predictive power. Statistical significance is of major importance in explanatory mod- eling, or profiling, but of secondary importance in predictive modeling (classification). In predictive modeling, statistically significant predictors might give hints as to more and less important predictors, but the even- tual choice of predictors should be based on predictive measures, such as the validation set confusion matrix (for classification) or the validation set lift chart (for ranking).

Appendix C: Logistic Regression for More Than Two Classes

The logistic model for a binary outcome can be extended for more than two classes. Suppose that there are m classes. Using a logistic regression model, for each record we would have m probabilities of belonging to each of the m classes. Since the m probabilities must add up to 1, we need estimate only m−1 probabilities.

Ordinal Classes Ordinal classes are classes that have a meaningful order. For example, in stock recommendations, the three classes buy, hold, and sell can be treated as ordered. As a simple rule, if classes can be numbered in a meaningful way, we consider them ordinal. When the number of classes is large (typically, more than 5), we can treat the outcome variable as continuous and perform multiple linear regression. When m = 2, the logistic model described above is used. We therefore need an extension of the logistic regression for a small number of ordinal classes (3 ≤ m ≤ 5). There are several ways to extend the binary-class case. Here, we describe the proportional odds or cumulative logit method. For other methods, see Hosmer and Lemeshow (2000).

For simplicity of interpretation and computation, we look at cumulative prob- abilities of class membership. For example, in the stock recommendations, we have m = 3 classes. Let us denote them by 1 = buy, 2 = hold, and 3 = sell. The probabilities estimated by the model are P (Y ≤ 1), (the probability of a buy recommendation) and P (Y ≤ 2) (the probability of a buy or hold recommen- dation). The three noncumulative probabilities of class membership can easily be recovered from the two cumulative probabilities:

P (Y = 1) = P (Y ≤ 1),

P (Y = 2) = P (Y ≤ 2)− P (Y ≤ 1),

P (Y = 3) = 1− P (Y ≤ 2).

APPENDIX: LOGISTIC REGRESSION FOR PROFILING 265

Next, we want to model each logit as a function of the predictors. Corre- sponding to each of them−1 cumulative probabilities is a logit. In our example, we would have

logit(buy) = log P (Y ≤ 1)

1− P (Y ≤ 1) ,

logit(buy or hold) = log P (Y ≤ 2)

1− P (Y ≤ 2) .

Each of the logits is then modeled as a linear function of the predictors (as in the two-class case). If in the stock recommendations we have a single predictor value x, we compute two logit values using two equations

logit(buy) = α0 + β1x,

logit(buy or hold) = β0 + β1x.

This means that both lines have the same slope (β1) but different intercepts. Once the coefficients α0, β0, β1 are estimated, we can compute the class membership probabilities by rewriting the logit equations in terms of probabilities. For the three-class case, for example, we would have

P (Y = 1) = P (Y ≤ 1) = 1 1 + e−(a0+b1x)

,

P (Y = 2) = P (Y ≤ 2)− P (Y ≤ 1) = 1 1 + e−(b0+b1x)

− 1 1 + e−(a0+b1x)

,

P (Y = 3) = 1− P (Y ≤ 2) = 1− 1 1 + e−(b0+b1x)

,

where a0, b0, and b1 are the estimates obtained from the training set. For each record, we now have the estimated probabilities that it belongs to

each of the classes. In our example, each stock would have three probabilities: for a buy recommendation, a hold recommendation, and a sell recommendation. The last step is to classify the record into one of the classes. This is done by assigning it to the class with the highest membership probability. For example, if a stock had estimated probabilities P (Y = 1) = 0.2, P (Y = 2) = 0.3, and P (Y = 3) = 0.5, we would classify it as getting a sell recommendation.

Nominal Classes When the classes cannot be ordered and are simply different from one another, we are in the case of nominal classes. An example is the choice between several brands of cereal. A simple way to verify that the classes are nominal is when it makes sense to tag them as A,B,C, . . ., and the assignment of letters to classes does not matter. For simplicity, let us assume that there are m = 3 brands of cereal that consumers can choose from (assuming that

266 LOGISTIC REGRESSION

each consumer chooses one). Then we estimate the probabilities P (Y = A), P (Y = B), and P (Y = C). As before, if we know two of the probabilities, the third probability is determined. We therefore use one of the classes as the reference class. Let us use C as the reference brand.

The goal, once again, is to model the class membership as a function of predictors. So in the cereals example, we might want to predict which cereal will be chosen if we know the cereal’s price x.

Next, we form m−1 pseudologit equations that are linear in the predictors. In our example, we would have

logit(A) = log P (Y = A)

P (Y = C) = α0 + α1x,

logit(B) = log P (Y = B)

P (Y = C) = β0 + β1x.

Once the four coefficients are estimated from the training set, we can estimate the class membership probabilities8:

P (Y = A) = ea0+a1x

1 + ea0+a1x + eb0+b1x ,

P (Y = B) = eb0+b1x

1 + ea0+a1x + eb0+b1x ,

P (Y = C) = 1− P (Y = A)− P (Y = B),

where a0, a1, b0, and b1 are the coefficient estimates obtained from the training set. Finally, a record is assigned to the class that has the highest probability.

Table 10.11 presents the R code for ordinal and nominal multinomial regression.

8From the two logit equations, we see that

P (Y = A) = P (Y = C)eα0+α1x

P (Y = B) = P (Y = C)eβ0+β1x

Since P (Y = A) + P (Y = B) + P (Y = C) = 1, we get

P (Y = C) = 1− P (Y = C)eα0+α1x − P (Y = C)eβ0+β1x

= 1

eα0+α1x+e β0+β1x

.

By plugging this form into the two equations above it, we also obtain the membership probabilities in classes A and B.

APPENDIX: LOGISTIC REGRESSION FOR PROFILING 267

TABLE 10.11 ORDINAL AND NOMINAL MULTINOMIAL REGRESSION IN R

code for logistic regression with more than 2 classes

# simulate simple data Y = rep(c("a", "b", "c"), 100) x = rep(c(1, 2, 3), 100) + rnorm(300, 0, 1)

# ordinal logistic regression library(MASS) Y = factor(Y, ordered = T) polr(Y ~ x)

# nominal logistic regression library(nnet) Y = factor(Y, ordered = F) multinom(Y ~ x)

Output

> polr(Y ~ x) Call: polr(formula = Y ~ x)

Coefficients: x

1.110152

Intercepts: a|b b|c

1.328406 3.218809

Residual Deviance: 543.3761 AIC: 549.3761

> multinom(Y ~ x) # weights: 9 (4 variable) initial value 329.583687 iter 10 value 268.610761 iter 10 value 268.610760 final value 268.610760 converged Call: multinom(formula = Y ~ x)

Coefficients: (Intercept) x

b -1.715653 1.058281 c -3.399917 1.738026

Residual Deviance: 537.2215 AIC: 545.2215

268 LOGISTIC REGRESSION

PROBLEMS

10.1 Financial Condition of Banks. The file Banks.csv includes data on a sample of 20 banks. The “Financial Condition” column records the judgment of an expert on the financial condition of each bank. This outcome variable takes one of two possible values—weak or strong—according to the financial condition of the bank. The predictors are two ratios used in the financial analysis of banks: TotLns&Lses/Assets is the ratio of total loans and leases to total assets and TotExp/Assets is the ratio of total expenses to total assets. The target is to use the two ratios for classifying the financial condition of a new bank.

Run a logistic regression model (on the entire dataset) that models the status of a bank as a function of the two financial measures provided. Specify the success class as weak (this is similar to creating a dummy that is 1 for financially weak banks and 0 otherwise), and use the default cutoff value of 0.5.

a. Write the estimated equation that associates the financial condition of a bank with its two predictors in three formats:

i. The logit as a function of the predictors

ii. The odds as a function of the predictors

iii. The probability as a function of the predictors

b. Consider a new bank whose total loans and leases/assets ratio = 0.6 and total expenses/assets ratio = 0.11. From your logistic regression model, estimate the following four quantities for this bank (use R to do all the intermediate calcula- tions; show your final answers to four decimal places): the logit, the odds, the probability of being financially weak, and the classification of the bank (use cutoff = 0.5).

c. The cutoff value of 0.5 is used in conjunction with the probability of being finan- cially weak. Compute the threshold that should be used if we want to make a classification based on the odds of being financially weak, and the threshold for the corresponding logit.

d. Interpret the estimated coefficient for the total loans & leases to total assets ratio (TotLns&Lses/Assets) in terms of the odds of being financially weak.

e. When a bank that is in poor financial condition is misclassified as financially strong, the misclassification cost is much higher than when a financially strong bank is misclassified as weak. To minimize the expected cost of misclassification, should the cutoff value for classification (which is currently at 0.5) be increased or decreased?

10.2 Identifying Good System Administrators. A management consultant is study- ing the roles played by experience and training in a system administrator’s ability to complete a set of tasks in a specified amount of time. In particular, she is interested in discriminating between administrators who are able to complete given tasks within a specified time and those who are not. Data are collected on the performance of 75 randomly selected administrators. They are stored in the file SystemAdministrators.csv.

The variable Experience measures months of full-time system administrator expe- rience, while Training measures the number of relevant training credits. The outcome variable Completed is either Yes or No, according to whether or not the administrator completed the tasks.

PROBLEMS 269

a. Create a scatter plot of Experience vs. Training using color or symbol to distinguish programmers who completed the task from those who did not complete it. Which predictor(s) appear(s) potentially useful for classifying task completion?

b. Run a logistic regression model with both predictors using the entire dataset as training data. Among those who completed the task, what is the percentage of programmers incorrectly classified as failing to complete the task?

c. To decrease the percentage in part (b), should the cutoff probability be increased or decreased?

d. How much experience must be accumulated by a programmer with 4 years of training before his or her estimated probability of completing the task exceeds 0.5?

10.3 Sales of Riding Mowers. A company that manufactures riding mowers wants to identify the best sales prospects for an intensive sales campaign. In particular, the man- ufacturer is interested in classifying households as prospective owners or nonowners on the basis of Income (in $1000s) and Lot Size (in 1000 ft2). The marketing expert looked at a random sample of 24 households, given in the file RidingMowers.csv. Use all the data to fit a logistic regression of ownership on the two predictors.

a. What percentage of households in the study were owners of a riding mower?

b. Create a scatter plot of Income vs. Lot Size using color or symbol to distinguish owners from nonowners. From the scatter plot, which class seems to have a higher average income, owners or nonowners?

c. Among nonowners, what is the percentage of households classified correctly?

d. To increase the percentage of correctly classified nonowners, should the cutoff probability be increased or decreased?

e. What are the odds that a household with a $60K income and a lot size of 20,000 ft2 is an owner?

f. What is the classification of a household with a $60K income and a lot size of 20,000 ft2? Use cutoff = 0.5.

g. What is the minimum income that a household with 16,000 ft2 lot size should have before it is classified as an owner?

10.4 Competitive Auctions on eBay.com. The file eBayAuctions.csv contains informa- tion on 1972 auctions transacted on eBay.com during May–June 2004. The goal is to use these data to build a model that will distinguish competitive auctions from non- competitive ones. A competitive auction is defined as an auction with at least two bids placed on the item being auctioned. The data include variables that describe the item (auction category), the seller (his or her eBay rating), and the auction terms that the seller selected (auction duration, opening price, currency, day of week of auction close). In addition, we have the price at which the auction closed. The goal is to predict whether or not an auction of interest will be competitive.

Data preprocessing. Create dummy variables for the categorical predictors. These include Category (18 categories), Currency (USD, GBP, Euro), EndDay (Monday–Sunday), and Duration (1, 3, 5, 7, or 10 days).

a. Create pivot tables for the mean of the binary outcome (Competitive?) as a function of the various categorical variables (use the original variables, not the dummies). Use the information in the tables to reduce the number of dummies that will be used in the model. For example, categories that appear most similar with respect to the distribution of competitive auctions could be combined.

270 LOGISTIC REGRESSION

b. Split the data into training (60%) and validation (40%) datasets. Run a logistic model with all predictors with a cutoff of 0.5.

c. If we want to predict at the start of an auction whether it will be competitive, we cannot use the information on the closing price. Run a logistic model with all predictors as above, excluding price. How does this model compare to the full model with respect to predictive accuracy?

d. Interpret the meaning of the coefficient for closing price. Does closing price have a practical significance? Is it statistically significant for predicting competitiveness of auctions? (Use a 10% significance level.)

e. Use stepwise selection (use function step() in the stats package or function stepAIC() in the MASS package) and an exhaustive search (use function glmulti() in package glmulti) to find the model with the best fit to the training data. Which predictors are used?

f. Use stepwise selection and an exhaustive search to find the model with the lowest predictive error rate (use the validation data). Which predictors are used?

g. What is the danger of using the best predictive model that you found?

h. Explain why the best-fitting model and the best predictive models are the same or different.

i. If the major objective is accurate classification, what cutoff value should be used?

j. Based on these data, what auction settings set by the seller (duration, opening price, ending day, currency) would you recommend as being most likely to lead to a competitive auction?

CHAPTER 11

Neural Nets

In this chapter, we describe neural networks, a flexible data-driven method that can be used for classification or prediction. Although considered a “blackbox” in terms of interpretability, neural nets have been highly successful in terms of pre- dictive accuracy. We discuss the concepts of “nodes” and “layers” (input layers, output layers, and hidden layers) and how they connect to form the structure of a network. We then explain how a neural network is fitted to data using a numerical example. Because overfitting is a major danger with neural nets, we present a strategy for avoiding it. We describe the different parameters that a user must specify and explain the effect of each on the process. Finally, we discuss the usefulness of neural nets and their limitations.

11.1 Introduction

Neural networks, also called artificial neural networks, are models for classification and prediction. The neural network is based on a model of biological activity in the brain, where neurons are interconnected and learn from experience. Neural networks mimic the way that human experts learn. The learning and memory properties of neural networks resemble the properties of human learning and memory, and they also have a capacity to generalize from particulars.

A number of successful applications have been reported in financial appli- cations (see Trippi and Turban, 1996) such as bankruptcy predictions, currency market trading, picking stocks and commodity trading, detecting fraud in credit card and monetary transactions, and customer relationship management (CRM). There have also been a number of successful applications of neural nets in

271

272 NEURAL NETS

engineering applications. One of the best known is ALVINN, an autonomous vehicle driving application for normal speeds on highways. Using as input a 30×32 grid of pixel intensities from a fixed camera on the vehicle, the classifier provides the direction of steering. The outcome variable is a categorical one with 30 classes, such as sharp left, straight ahead, and bear right.

The main strength of neural networks is their high predictive performance. Their structure supports capturing very complex relationships between predic- tors and an outcome variable, which is often not possible with other predictive models.

11.2 Concept and Structure of a Neural Network

The idea behind neural networks is to combine the predictor information in a very flexible way that captures complicated relationships among these variables and between them and the outcome variable. For instance, recall that in linear regression models, the form of the relationship between the outcome and the predictors is specified directly by the user (see Chapter 6.) In many cases, the exact form of the relationship is very complicated, or is generally unknown. In linear regression modeling we might try different transformations of the predic- tors, interactions between predictors, and so on, but the specified form of the relationship remains linear. In comparison, in neural networks the user is not required to specify the correct form. Instead, the network tries to learn about such relationships from the data. In fact, linear regression and logistic regression can be thought of as special cases of very simple neural networks that have only input and output layers and no hidden layers.

Although researchers have studied numerous different neural network archi- tectures, the most successful applications of neural networks in data mining have been multilayer feedforward networks. These are networks in which there is an input layer consisting of nodes (sometimes called neurons) that simply accept the pre- dictor values, and successive layers of nodes that receive input from the previous layers. The outputs of nodes in each layer are inputs to nodes in the next layer. The last layer is called the output layer. Layers between the input and output layers are known as hidden layers. A feedforward network is a fully connected network with a one-way flow and no cycles. Figure 11.1 shows a diagram for this architecture, with two hidden layers and one node in the output layer rep- resenting the outcome value to be predicted. In a classification problem with m classes, there would be m output nodes (or m− 1 output nodes, depending on the software).

FITTING A NETWORK TO DATA 273

FIGURE 11.1 MULTILAYER FEEDFORWARD NEURAL NETWORK

11.3 Fitting a Network to Data

To illustrate how a neural network is fitted to data, we start with a very small illustrative example. Although the method is by no means operational in such a small example, it is useful for explaining the main steps and operations, for showing how computations are done, and for integrating all the different aspects of neural network data fitting. We will later discuss a more realistic setting.

Example 1: Tiny Dataset

Consider the following very small dataset. Table 11.1 includes information on a tasting score for a certain processed cheese. The two predictors are scores for fat and salt, indicating the relative presence of fat and salt in the particular cheese sample (where 0 is the minimum amount possible in the manufacturing process, and 1 the maximum). The outcome variable is the cheese sample’s consumer taste preference, where like or dislike indicate whether the consumer likes the cheese or not.

Figure 11.2 describes an example of a typical neural net that could be used for predicting cheese preference (like/dislike) by new consumers, based on these data. We numbered the nodes in the example from 1 to 7. Nodes 1 and 2 belong to the input layer, nodes 3 to 5 belong to the hidden layer, and nodes 6 and 7 belong to the output layer. The values on the connecting arrows are called weights, and the weight on the arrow from node i to node j is denoted by wi,j . The additional bias nodes, denoted by θj , serve as an intercept for the output from node j. These are all explained in further detail below.

274 NEURAL NETS

TABLE 11.1 TINY EXAMPLE ON TASTING SCORES FOR SIX CONSUMERS AND TWO PREDICTORS

Obs. Fat Score Salt Score Acceptance

1 0.2 0.9 like 2 0.1 0.1 dislike 3 0.2 0.4 dislike 4 0.2 0.5 dislike 5 0.4 0.5 like 6 0.3 0.8 like

Computing Output of Nodes

We discuss the input and output of the nodes separately for each of the three types of layers (input, hidden, and output). The main difference is the function used to map from the input to the output of the node.

Input nodes take as input the values of the predictors. Their output is the same as the input. If we have p predictors, the input layer will usually include p nodes. In our example, there are two predictors, and therefore the input layer (shown in Figure 11.2) includes two nodes, each feeding into each node of the hidden layer. Consider the first record: The input into the input layer is Fat = 0.2 and Salt = 0.9, and the output of this layer is also x1 = 0.2 and x2 = 0.9.

Hidden layer nodes take as input the output values from the input layer. The hidden layer in this example consists of three nodes, each receiving input from all the input nodes. To compute the output of a hidden layer node, we compute a weighted sum of the inputs and apply a certain function to it. More formally, for a set of input values x1, x2, . . . , xp, we compute the output of

1

2

3

4

5

6

Input Layer Hidden Layer Output Layer

w13

Fat

Salt

w14

w23

w25

w24

w15

w36

w46

w56

θ3

θ4

θ5

θ6

dislike

like7

w47

w37

w57

θ 7

FIGURE 11.2 NEURAL NETWORK FOR THE TINY EXAMPLE. CIRCLES REPRESENT NODES (“NEURONS”), Wi,j ON ARROWS ARE WEIGHTS, AND θj ARE NODE BIAS VALUES

FITTING A NETWORK TO DATA 275

node j by taking the weighted sum1 θj + ∑p

i=1 wijxi, where θj, w1,j, . . . , wp,j are weights that are initially set randomly, then adjusted as the network “learns.” Note that θj , also called the bias of node j, is a constant that controls the level of contribution of node j. In the next step, we take a function g of this sum. The function g, also called a transfer function or activation function is some monotone function. Examples include the linear function [g(s) = bs], an exponential function [g(s) = exp(bs)], and a logistic/sigmoidal function [g(s) = 1/1 + e−s]. This last function is by far the most popular one in neural networks. Its practical value arises from the fact that it has a squashing effect on very small or very large values but is almost linear in the range where the value of the function is between 0.1 and 0.9.

If we use a logistic activation function, we can write the output of node j in the hidden layer as

Outputj = g

( θj +

p∑ i=1

wijxi

) =

1

1 + e−(θj+ ∑p

i=1 wijxi) . (11.1)

Initializing the Weights The values of θj and wij are initialized to small, usually random, numbers (typically, but not always, in the range 0.00 ± 0.05). Such values represent a state of no knowledge by the network, similar to a model with no predictors. The initial weights are used in the first round of training.

Returning to our example, suppose that the initial weights for node 3 are θ3 = −0.3, w1,3 = 0.05, and w2,3 = 0.01 (as shown in Figure 11.3). Using

1

2

3

4

5

Input Layer Hidden Layer Output Layer

0.05

–0.01

0.2 0.030.01

–0.01

0.01

0.05

0.015

Fat = 0.2

Salt = 0.9

0.2

0.9

0.43

0.51

0.52

0.506

–0.3

0.02

0.05

–0.015

dislike

like

6

7

–0.02

–0.03

–0.02

–0.04

0.481

FIGURE 11.3 COMPUTING NODE OUTPUTS (IN BOLDFACE TYPE) USING THE FIRST RECORD IN THE TINY EXAMPLE AND A LOGISTIC FUNCTION

1Other options exist for combining inputs, such as taking the maximum or minimum of the weighted inputs rather than their sum, but they are much less popular.

276 NEURAL NETS

the logistic function, we can compute the output of node 3 in the hidden layer (using the first record) as

Output3 = 1

1 + e−[−0.3+(0.05)(0.2)+(0.01)(0.9)] = 0.43.

Figure 11.3 shows the initial weights, inputs, and outputs for the first record in our tiny example. If there is more than one hidden layer, the same calculation applies, except that the input values for the second, third, and so on, hidden layers would be the output of the preceding hidden layer. This means that the number of input values into a certain node is equal to the number of nodes in the preceding layer. (If there was an additional hidden layer in our example, its nodes would receive input from the three nodes in the first hidden layer.)

Finally, the output layer obtains input values from the (last) hidden layer. It applies the same function as above to create the output. In other words, it takes a weighted sum of its input values and then applies the function g. In our example, output nodes 6 and 7 receive input from the three hidden layer nodes. We can compute the output of these nodes by

Output6 = 1

1 + e−[−0.04+ (−0.02)(0.43)+ (−0.03)(0.51)+ (0.015)(0.52)] = 0.481

Output7 = 1

1 + e−[−0.015+ (0.01)(0.430)+ (0.05)(0.507)+ (0.015)(0.511)] = 0.506.

These two numbers are almost the propensities P(Y = dislike | Fat = 0.2, Salt = 0.9) and P(Y = like | Fat = 0.2. Salt = 0.9). The last step involves normalizing these two values so that they add up to 1. In other words,

P(Y = dislike) = Output6/(Output6 + Output7) = 0.481/(0.481 + 0.506) = 0.49

P(Y = like) = 1 − P(Y = dislike) = 0.506/(0.481 + 0.506) = 0.51

For classification, we use a cutoff value (for a binary outcome) on the propensity. Using a cutoff of 0.5, we would classify this record as like. For applications with more than two classes, we choose the output node with the largest value.

Relation to Linear and Logistic Regression Consider a neural net- work with a single output node and no hidden layers. For a dataset with p predictors, the output node receives x1, x2, . . . , xp, takes a weighted sum of these, and applies the g function. The output of the neural network is therefore g (θ +

∑p i=1 wixi).

First, consider a numerical outcome variable Y . If g is the identity function [g(s) = s], the output is simply

Ŷ = θ +

p∑ i=1

wixi.

FITTING A NETWORK TO DATA 277

This is exactly equivalent to the formulation of a multiple linear regression! This means that a neural network with no hidden layers, a single output node, and an identity function g searches only for linear relationships between the outcome and the predictors.

Now consider a binary output variable Y . If g is the logistic function, the output is simply

P̂ (Y = 1) = 1

1 + e−(θ+ ∑p

i=1 wixi) ,

which is equivalent to the logistic regression formulation! In both cases, although the formulation is equivalent to the linear and logistic

regression models, the resulting estimates for the weights (coefficients in linear and logistic regression) can differ, because the estimation method is different. The neural net estimation method is different from least squares, the method used to calculate coefficients in linear regression, or the maximum likelihood method used in logistic regression. We explain below the method by which the neural network learns.

Preprocessing the Data

When using a logistic activation function (option act.fct = ’logistic’ in R), neural networks perform best when the predictors and outcome variable are on a scale of [0,1]. For this reason, all variables should be scaled to a [0,1] interval before entering them into the network. For a numerical variable X that takes values in the range [a, b] where a < b, we normalize the measurements by subtracting a and dividing by b− a. The normalized measurement is then

Xnorm = X − a b− a

.

Note that if [a, b] is within the [0,1] interval, the original scale will be stretched. If a and b are unknown, we can estimate them from the minimal and maximal

values of X in the data. Even if new data exceed this range by a small amount, yielding normalized values slightly lower than 0 or larger than 1, this will not affect the results much.

For binary variables, no adjustment is needed other than creating dummy variables. For categorical variables with m categories, if they are ordinal in nature, a choice of m fractions in [0,1] should reflect their perceived ordering. For example, if four ordinal categories are equally distant from each other, we can map them to [0, 0.25, 0.5, 1]. If the categories are nominal, transforming into m− 1 dummies is a good solution.

Another operation that improves the performance of the network is to trans- form highly skewed predictors. In business applications, there tend to be many highly right-skewed variables (such as income). Taking a log transform of a

278 NEURAL NETS

right-skewed variable (before converting to a [0,1] scale) will usually spread out the values more symmetrically.

Another common sigmoidal function is the hyperbolic tangent (option act.fct = ’tanh’ in R). When using this function, it is usually better to scale predictors to a [-1,1] scale.

Training the Model

Training the model means estimating the weights θj and wij that lead to the best predictive results. The process that we described earlier (Section 11.1) for computing the neural network output for a record is repeated for all records in the training set. For each record, the model produces a prediction which is then compared with the actual outcome value. Their difference is the error for the output node. However, unlike least squares or maximum likelihood, where a global function of the errors (e.g., sum of squared errors) is used for estimating the coefficients, in neural networks, the estimation process uses the errors iteratively to update the estimated weights.

In particular, the error for the output node is distributed across all the hidden nodes that led to it, so that each node is assigned “responsibility” for part of the error. Each of these node-specific errors is then used for updating the weights.

Back Propagation of Error The most popular method for using model errors to update weights (“learning”) is an algorithm called back propagation. As the name implies, errors are computed from the last layer (the output layer) back to the hidden layers.

Let us denote by ŷk the output from output node k. The error associated with output node k is computed by

errk = ŷk(1− ŷk)(yk − ŷk).

Notice that this is similar to the ordinary definition of an error (yk − ŷk) multi- plied by a correction factor. The weights are then updated as follows:

θnewj = θ old j + l × errj, (11.2)

wnewi,j = w old i,j + l × errj,

where l is a learning rate or weight decay parameter, a constant ranging typically between 0 and 1, which controls the amount of change in weights from one iteration to the next.

In our example, the error associated with output node 7 for the first record is (0.506)(1 − 0.506)(1 − 0.506) = 0.123. For output node 6 the error is 0.481(1− 0.481)(1− 0.481) = 0.129. These errors are then used to compute the errors associated with the hidden layer nodes, and those weights are updated accordingly using a formula similar to (11.2).

FITTING A NETWORK TO DATA 279

Two methods for updating the weights are case updating and batch updating. In case updating, the weights are updated after each record is run through the network (called a trial). For example, if we used case updating in the tiny exam- ple, the weights would first be updated after running record 1 as follows: Using a learning rate of 0.5, the weights θ7, w3,7, w4,7, and w5,7 are updated to

θ7 = −0.015 + (0.5)(0.123) = 0.047 w3,7 = 0.01 + (0.5)(0.123) = 0.072

w4,7 = 0.05 + (0.5)(0.123) = 0.112

w5,7 = 0.015 + (0.5)(0.123) = 0.077

Similarly, we obtain updated weights θ6 = 0.025, w3,6 = 0.045, w4,6 = 0.035, and w5,6 = 0.045. These new weights are next updated after the second record is run through the network, the third, and so on, until all records are used. This is called one epoch, sweep, or iteration through the data. Typically, there are many iterations.

In batch updating, the entire training set is run through the network before each updating of weights takes place. In that case, the errors errk in the updating equation is the sum of the errors from all records. In practice, case updating tends to yield more accurate results than batch updating, but requires a longer run time. This is a serious consideration, since even in batch updating, hundreds or even thousands of sweeps through the training data are executed.

When does the updating stop? The most common conditions are one of the following:

1. When the new weights are only incrementally different from those of the preceding iteration

2. When the misclassification rate reaches a required threshold

3. When the limit on the number of runs is reached

Let us examine the output from running a neural network on the tiny data. Following Figures 11.2 and 11.3, we used a single hidden layer with three nodes. R has several packages for neural nets, the most common ones are nnet and neuralnet (note that package nnet does not enable multilayer networks and has no plotting option). We used neuralnet in this example. The weights and model output are shown in Table 11.2. Figure 11.4 shows these weights in a format similar to that of our previous diagrams.

The first 3×3 table shows the weights that connect the input layer and the hidden layer. The Bias nodes (first row of the weights table) are the weights θ3, θ4, and θ5. The weights in this table are used to compute the output of the hidden layer nodes. They were computed iteratively after choosing a random initial set of weights (like the set we chose in Figure 11.3). We use the weights in the way

280 NEURAL NETS

TABLE 11.2 NEURAL NETWORK WITH A SINGLE HIDDEN LAYER (THREE NODES) FOR THE TINY DATA EXAMPLE

code for running a neural network

nn <- neuralnet(Like + Dislike ~ Salt + Fat, data = df, linear.output = F, hidden = 3)

# display weights nn$weights

# display predictions prediction(nn)

# plot network plot(nn, rep="best")

Output

> nn$weights [[1]] [[1]][[1]]

[,1] [,2] [,3] [1,] -1.061143694 3.057021840 3.337952001 [2,] 2.326024132 -3.408663181 -4.293213530 [3,] 4.106434697 -6.525668384 -5.929418648

[[1]][[2]] [,1] [,2]

[1,] -0.3495332882 -1.677855862 [2,] 5.8777145665 -3.606625360 [3,] -5.3529200726 5.620329700 [4,] -6.1115038896 6.696286857

> prediction(nn) Data Error: 0; $rep1

Salt Fat Like Dislike 1 0.1 0.1 0.0002415535993 0.99965512479 2 0.4 0.2 0.0344215786564 0.96556787694 3 0.5 0.2 0.1248666747740 0.87816827940 4 0.9 0.2 0.9349452648141 0.07022732257 5 0.8 0.3 0.9591361793188 0.04505630529 6 0.5 0.4 0.8841904620140 0.12672437721

−5.92942

−6 .52

56 7

4. 10

64 3

Fat

−4.29321

−3.40866

2.32 602

Salt

6.69 629−6

.1 11

5

5.62033

−5 .35

29 2

−3.60663

5.87771

Dislike

Like

3.33795 3.05702

−1.06114

1

−1.67786

−0.34953

1

Error: 0.03757 Steps: 76

FIGURE 11.4 NEURAL NETWORK FOR THE TINY EXAMPLE WITH FINAL WEIGHTS FROM R OUTPUT. VALUES ON THE FIVE DOWNWARDS ARROWS DENOTE THE BIAS

FITTING A NETWORK TO DATA 281

we described earlier to compute the hidden layer’s output. For instance, for the first record, the output of our previous node 3 is:

Output3 = 1

1 + e−[−1.06+(2.33)(0.9)+(4.11)(0.2)] = 0.86.

Similarly, we can compute the output from the two other hidden nodes for the same record and get Output4 = 0.21 and Output5 = 0.15. The second 4×2 table gives the weights connecting the hidden and output layer nodes. To compute the probability (= propensity) for the dislike output node for the first record, we use the outputs from the hidden layer that we computed above, and get

Outputdislike = 1

1 + e−[−1.68+ (−3.61)(0.86)+ (5.62)(0.21)+ (6.70)(0.15)] = 0.07.

Similarly, we can compute the probability for the like output node, obtaining the value Output7 = 0.93.

The probabilities for the other five records are computed in the same manner, replacing the input value in the computation of the hidden layer outputs and then plugging these outputs into the computation for the output layer. The confusion matrix based on these probabilities, using and a cutoff of 0.5, is given in Table 11.3. We can see that the network correctly classifies all six records.

TABLE 11.3 CONFUSION MATRIX FOR THE TINY EXAMPLE

code for the confusion matrix

library(caret) predict <- compute(nn, data.frame(df$Salt, df$Fat)) predicted.class=apply(predict$net.result,1,which.max)-1 confusionMatrix(ifelse(predicted.class=="1", "dislike", "like"), df$Acceptance)

Output

> confusionMatrix(ifelse(predicted.class=="1", "dislike", "like"), + df$Acceptance) Confusion Matrix and Statistics

Reference Prediction dislike like

dislike 3 0 like 0 3

Accuracy : 1

282 NEURAL NETS

Example 2: Classifying Accident Severity

Let’s apply the network training process to some real data: US automobile acci- dents that have been classified by their level of severity as no injury, injury, or fatality. A firm might be interested in developing a system for quickly classifying the severity of an accident, based on initial reports and associated data in the system (some of which rely on GPS-assisted reporting). Such a system could be used to assign emergency response team priorities. Table 11.4 shows a small extract (10 records, four predictor variables) from a US government database.

The explanation of the four predictor variables and outcome variable is given in Table 11.5. For the analysis, we converted ALCHL_I to a 0/1 dummy variable (1 = presence of alcohol) and created four dummies for SUR_COND. This gives us a total of seven predictors.

With the exception of alcohol involvement and a few other variables in the larger database, most of the variables are ones that we might reasonably expect to be available at the time of the initial accident report, before accident details and severity have been determined by first responders. A data mining model that could predict accident severity on the basis of these initial reports would have value in allocating first responder resources.

To use a neural net architecture for this classification problem, we use seven nodes in the input layer, one for each of the seven predictors, and three neurons

TABLE 11.4 SUBSET FROM THE ACCIDENTS DATA, FOR A HIGH- FATALITY REGION

Obs. ALCHL_I PROFIL_I_R SUR_COND VEH_INVL MAX_SEV_IR

1 1 1 1 1 1 2 2 1 1 1 0 3 2 1 1 1 1 4 1 1 1 1 0 5 2 1 1 1 2 6 2 0 1 1 1 7 2 0 1 3 1 8 2 0 1 4 1 9 2 0 1 2 0 10 2 0 1 2 0

TABLE 11.5 DESCRIPTION OF VARIABLES FOR AUTOMOBILE ACCIDENT EXAMPLE

ALCHL_I Presence (1) or absence (2) of alcohol PROFIL_I_R Profile of the roadway: level (1), other (0) SUR_COND Surface condition of the road: dry (1), wet (2), snow/slush (3), ice (4), unknown (9) VEH_INVL Number of vehicles involved MAX_SEV_IR Presence of injuries/fatalities: no injuries (0), injury (1), fatality (2)

FITTING A NETWORK TO DATA 283

(one for each class) in the output layer. We use a single hidden layer and experi- ment with the number of nodes. If we increase the number of nodes from one to five and examine the resulting confusion matrices, we find that two nodes gives a good balance between improving the predictive performance on the training set without deteriorating the performance on the validation set. (Networks with more than two nodes in the hidden layer performed as well as the two-node net- work, but add undesirable complexity.) Tables 11.6 and 11.7 show the R code and output for the accidents data.

Our results can depend on how we set the different parameters, and there are a few pitfalls to avoid. We discuss these next.

Avoiding Overfitting

A weakness of the neural network is that it can easily overfit the data, causing the error rate on validation data (and most important, on new data) to be too large. It is therefore important to limit the number of training iterations and not to over-train on the data (e.g., in R’s neuralnet() function you can control the number of iterations using argument stepmax). As in classification and regression trees, overfitting can be detected by examining the performance on the validation set, or better, on a cross-validation set, and seeing when it starts deteriorating, while the training set performance is still improving. This approach is used in some algorithms, to limit the number of training iterations. The validation error decreases in the early iterations of the training but after a while, it begins to increase. The point of minimum validation error is a good indicator of the best number of iterations for training, and the weights at that stage are likely to provide the best error rate in new data.

Using the Output for Prediction and Classification

When the neural network is used for predicting a numerical outcome variable, the resulting output needs to be scaled back to the original units of that outcome variable. Recall that numerical variables (both predictor and outcome variables) are usually rescaled to a [0,1] interval before being used by the network. The output will therefore also be on a [0,1] scale. To transform the prediction back to the original y units, which were in the range [a, b], we multiply the network output by b− a and add a.

When the neural net is used for classification and we have m classes, we will obtain an output from each of the m output nodes (or m − 1 output nodes depending on the software.2 How do we translate these m outputs into a

2Function neuralnet() by default generated m− 1 output nodes for an outcome with m classes. To get m output nodes, specify each class on the left-hand-side of the formula—see Table 11.2)

284 NEURAL NETS

TABLE 11.6 A NEURAL NETWORK WITH TWO NODES IN THE HIDDEN LAYER (ACCIDENTS DATA)

code for running and evaluating a neural net on the accidents data

library(neuralnet,nnet,caret)

accidents.df <- read.csv("Accidents.csv") # selected variables vars=c("ALCHL_I", "PROFIL_I_R", "VEH_INVL")

# partition the data set.seed(2) training=sample(row.names(accidents.df), dim(accidents.df)[1]*0.6) validation=setdiff(row.names(accidents.df), training)

# when y has multiple classes - need to dummify trainData <- cbind(accidents.df[training,c(vars)],

class.ind(accidents.df[training,]$SUR_COND), class.ind(accidents.df[training,]$MAX_SEV_IR))

names(trainData)=c(vars, paste("SUR_COND_", c(1, 2, 3, 4, 9), sep=""), paste("MAX_SEV_IR_", c(0, 1, 2), sep=""))

validData <- cbind(accidents.df[validation,c(vars)], class.ind(accidents.df[validation,]$SUR_COND), class.ind(accidents.df[validation,]$MAX_SEV_IR))

names(validData)=c(vars, paste("SUR_COND_", c(1, 2, 3, 4, 9), sep=""), paste("MAX_SEV_IR_", c(0, 1, 2), sep=""))

# run nn with 2 hidden nodes # use hidden= with a vector of integers specifying number of hidden nodes in each layer nn <- neuralnet(MAX_SEV_IR_0 + MAX_SEV_IR_1 + MAX_SEV_IR_2 ~

ALCHL_I + PROFIL_I_R + VEH_INVL + SUR_COND_1 + SUR_COND_2 + SUR_COND_3 + SUR_COND_4, data = trainData, hidden = 2)

training.prediction=compute(nn, trainData[,-c(8:11)]) training.class=apply(training.prediction$net.result,1,which.max)-1 confusionMatrix(training.class, accidents.df[training,]$MAX_SEV_IR)

validation.prediction=compute(nn, validData[,-c(8:11)]) validation.class=apply(validation.prediction$net.result,1,which.max)-1 confusionMatrix(validation.class, accidents.df[validation,]$MAX_SEV_IR)

Output

> confusionMatrix(training.class, accidents.df[training,]$MAX_SEV_IR) Confusion Matrix and Statistics

Reference Prediction 0 1 2

0 332 0 34 1 0 167 40 2 1 7 18

> confusionMatrix(validation.class, accidents.df[validation,]$MAX_SEV_IR) Confusion Matrix and Statistics

Reference Prediction 0 1 2

0 217 0 20 1 0 121 22 2 1 4 15

Overall Statistics

Accuracy : 0.8825

REQUIRED USER INPUT 285

TABLE 11.7 OUTPUT FOR NEURAL NETWORK FOR ACCIDENT DATA, WITH TWO NODES IN THE HIDDEN LAYER

> confusionMatrix(training.class, + accidents.df[training,]$MAX_SEV_IR) Confusion Matrix and Statistics

Reference Prediction 0 1 2

0 332 0 34 1 0 167 40 2 1 7 18

> confusionMatrix(validation.class, + accidents.df[validation,]$MAX_SEV_IR) Confusion Matrix and Statistics

Reference Prediction 0 1 2

0 217 0 20 1 0 121 22 2 1 4 15

Overall Statistics

Accuracy : 0.8825

classification rule? Usually, the output node with the largest value determines the net’s classification.

In the case of a binary outcome (m = 2), we use two output nodes with a cutoff value to map a predicted probability to one of the two classes. Although we typically use a cutoff of 0.5 with other classifiers, in neural networks there is a tendency for values to cluster around 0.5 (from above and below). An alter- native is to use the validation set to determine a cutoff that produces reasonable predictive performance.

11.4 Required User Input

One of the time-consuming and complex aspects of training a model using back propagation is that we first need to decide on a network architecture. This means specifying the number of hidden layers and the number of nodes in each layer. The usual procedure is to make intelligent guesses using past experience and to do several trial-and-error runs on different architectures. Algorithms exist that grow the number of nodes selectively during training or trim them in a manner analogous to what is done in classification and regression trees (see Chapter 9). Research continues on such methods. As of now, no automatic method seems clearly superior to the trial-and-error approach. A few general guidelines for choosing an architecture follow.

286 NEURAL NETS

Number of hidden layers. The most popular choice for the number of hidden layers is one. A single hidden layer is usually sufficient to capture even very complex relationships between the predictors.

Size of hidden layer. The number of nodes in the hidden layer also determines the level of complexity of the relationship between the predictors that the network captures. The trade-off is between under- and overfitting. On the one hand, using too few nodes might not be sufficient to capture complex relationships (recall the special cases of a linear relationship such as in linear and logistic regression, in the extreme case of zero nodes or no hidden layer). On the other hand, too many nodes might lead to overfitting. A rule of thumb is to start with p (number of predictors) nodes and gradually decrease or increase while checking for overfitting. The number of hidden layers and the size of each hidden layer can be specified in R’s neuralnet() function using argument hidden.

Number of output nodes. For a categorical outcome withm classes, the number of nodes should equal m or m − 1. For a numerical outcome, typically a single output node is used unless we are interested in predicting more than one function.

In addition to the choice of architecture, the user should pay attention to the choice of predictors. Since neural networks are highly dependent on the quality of their input, the choice of predictors should be done carefully, using domain knowledge, variable selection, and dimension reduction techniques before using the network. We return to this point in the discussion of advantages and weaknesses.

Depending on the software, other parameters that the user might be able to control are the learning rate (a.k.a. weight decay), l, and the momentum. The first is used primarily to avoid overfitting, by down-weighting new information. This helps to tone down the effect of outliers on the weights and avoids getting stuck in local optima. This parameter typically takes a value in the range [0, 1]. Berry and Linoff (2000) suggest starting with a large value (moving away from the random initial weights, thereby “learning quickly” from the data) and then slowly decreasing it as the iterations progress and as the weights are more reliable. Han and Kamber (2001) suggest the more concrete rule of thumb of setting l = 1/(current number of iterations). This means that at the start, l = 1, during the second iteration it is l = 0.5, and then it keeps decaying toward l = 0. In R, you can set the learning rate in function neuralnet() using argument learningrate.

The second parameter, called momentum, is used to “keep the ball rolling” (hence the term momentum) in the convergence of the weights to the optimum. The idea is to keep the weights changing in the same direction as they did in the preceding iteration. This helps to avoid getting stuck in a local optimum.

EXPLORING THE RELATIONSHIP BETWEEN PREDICTORS AND OUTCOME 287

High values of momentum mean that the network will be “reluctant” to learn from data that want to change the direction of the weights, especially when we consider case updating. In general, values in the range 0–2 are used.

11.5 Exploring the Relationship Between Predictors and Outcome

Neural networks are known to be “black boxes” in the sense that their output does not shed light on the patterns in the data that it models (like our brains). In fact, that is one of the biggest criticisms of the method. However, in some cases, it is possible to learn more about the relationships that the network captures by conducting a sensitivity analysis on the validation set. This is done by setting all predictor values to their mean and obtaining the network’s prediction. Then, the process is repeated by setting each predictor sequentially to its minimum, and then maximum, value. By comparing the predictions from different levels of the predictors, we can get a sense of which predictors affect predictions more and in what way.

U N S U P E R V I S E D F E A T U R E E X T R A C T I O N A N D D E E P L E A R N I N G

Most of the data we have dealt with in this book has been either numeric or cat- egorical, and the available predictor variables or features have been inherently informative (e.g., the size of a car’s engine, or a car’s gas mileage). It has been relatively straightforward to discover their impact on the target variable, and focus on the most meaningful ones. With some data, however, we begin with a huge mass of values that are not “predictors” in the same sense. With image data, one “variable” would be the color and intensity of the pixel in a particular position, and one 2-inch square image might have 40,000 pixels. We are now in the realm of “big data.”

How can we derive meaningful higher-level features such as edges, curves, shapes, and even higher-level features such as faces? Deep learning has made big strides in this area. Deep learning networks (DLNs) refer to neural nets with many hidden layers used to self-learn features from complex data. DLNs are especially effective at capturing local structure and dependencies in complex data, such as in images or audio. This is done by using neural nets in an unsupervised way, where the data are used as both the input (with some added noise) and the output. When the multiple hidden layers are used, the DLN’s learning is hierarchical---typically, from a single pixel to edges, from edges to higher-level features, and so on. The network then “pools” similar high-level features across images to build up clusters of similar features. If they appear often enough in large numbers of images, distinctive and regular structures like faces, vehicles, and houses will emerge as highest-level features. See Le et al. (2012) for a technical description of how this method yielded

288 NEURAL NETS

labels not just of faces but of different types of faces (e.g., cat faces, of which there are many on the Internet).

Other applications include speech and handwriting recognition. Facebook and Google are effectively using deep learning to identify faces from images, and high- end retail stores use facial recognition technology involving deep learning to iden- tify VIP customers and give them special sales attention.3

While DLNs are not new, they have only gained momentum recently, thanks to dramatic improvements in computing power, allowing us to deal both with the huge amount of data and the extreme complexity of the networks. The abundance of unlabeled data such as images and audio has also helped tremendously. As with neural networks, DLNs suffer from overfitting and slow runtime. Solutions include using regularization (removing “useless” nodes by zeroing out their coefficients) and tricks to improve over the back-propagation algorithm.

11.6 Advantages and Weaknesses of Neural Networks

The most prominent advantage of neural networks is their good predictive per- formance. They are known to have high tolerance to noisy data and the ability to capture highly complicated relationships between the predictors and an out- come variable. Their weakest point is in providing insight into the structure of the relationship, hence their blackbox reputation.

Several considerations and dangers should be kept in mind when using neural networks. First, although they are capable of generalizing from a set of examples, extrapolation is still a serious danger. If the network sees only records in a certain range, its predictions outside this range can be completely invalid.

Second, neural networks do not have a built-in variable selection mecha- nism. This means that there is a need for careful consideration of predictors. Combination with classification and regression trees (see Chapter 9) and other dimension reduction techniques (e.g., principal components analysis in Chapter 4) is often used to identify key predictors.

Third, the extreme flexibility of the neural network relies heavily on hav- ing sufficient data for training purposes. A related issue is that in classification problems, the network requires sufficient records of the minority class in order to learn it. This is achieved by oversampling, as explained in Chapter 2.

Fourth, a technical problem is the risk of obtaining weights that lead to a local optimum rather than the global optimum, in the sense that the weights converge

3www.npr.org/sections/alltechconsidered/2013/07/21/203273764/high-end-stores-use-facial- recognition-tools-to-spot-vips, accessed July 21, 2013.

ADVANTAGES AND WEAKNESSES OF NEURAL NETWORKS 289

to values that do not provide the best fit to the training data. We described several parameters used to try to avoid this situation (such as controlling the learning rate and slowly reducing the momentum). However, there is no guarantee that the resulting weights are indeed the optimal ones.

Finally, a practical consideration that can determine the usefulness of a neural network is the computation time. Neural networks are relatively heavy on com- putation time, requiring a longer runtime than other classifiers. This runtime grows greatly when the number of predictors is increased (as there will be many more weights to compute). In applications where real-time or near-real-time prediction is required, runtime should be measured to make sure that it does not cause unacceptable delay in the decision-making.

290 NEURAL NETS

PROBLEMS

11.1 Credit Card Use. Consider the hypothetical bank data in Table 11.8 on consumers’ use of credit card credit facilities. Create a small worksheet in Excel, like that used in Example 1, to illustrate one pass through a simple neural network.

TABLE 11.8 DATA FOR CREDIT CARD EXAMPLE AND VARIABLE DESCRIPTIONS

Years Salary Used Credit

4 43 0 18 65 1 1 53 0 3 95 0

15 88 1 6 112 1

Years: number of years the customer has been with the bank Salary: customer’s salary (in thousands of dollars) Used Credit: 1 = customer has left an unpaid credit card balance at the end of at least one month in the prior year, 0 = balance was paid off at the end of each month

11.2 Neural Net Evolution. A neural net typically starts out with random coefficients; hence, it produces essentially random predictions when presented with its first case. What is the key ingredient by which the net evolves to produce a more accurate prediction?

11.3 Car Sales. Consider the data on used cars (ToyotaCorolla.csv) with 1436 records and details on 38 attributes, including Price, Age, KM, HP, and other specifications. The goal is to predict the price of a used Toyota Corolla based on its specifications.

a. Fit a neural network model to the data. Use a single hidden layer with 2 nodes.

• Use predictors Age_08_04, KM, Fuel_Type, HP, Automatic, Doors, Quarterly_Tax, Mfr_Guarantee, Guarantee_Period, Airco, Automatic_airco, CD_Player, Powered_Windows, Sport_Model, and Tow_Bar.

• Remember to first scale the numerical predictor and outcome variables to a 0–1 scale (use function preprocess() with method = “range”—see Chapter 7) and convert categorical predictors to dummies.

Record the RMS error for the training data and the validation data. Repeat the process, changing the number of hidden layers and nodes to {single layer with 5 nodes}, {two layers, 5 nodes in each layer}.

i. What happens to the RMS error for the training data as the number of layers and nodes increases?

ii. What happens to the RMS error for the validation data?

iii. Comment on the appropriate number of layers and nodes for this application.

11.4 Direct Mailing to Airline Customers. East-West Airlines has entered into a part- nership with the wireless phone company Telcon to sell the latter’s service via direct mail. The file EastWestAirlinesNN.csv contains a subset of a data sample of who has already received a test offer. About 13% accepted.

PROBLEMS 291

You are asked to develop a model to classify East–West customers as to whether they purchase a wireless phone service contract (outcome variable Phone_Sale). This model will be used to classify additional customers.

a. Run a neural net model on these data, using a single hidden layer with 5 nodes. Remember to first convert categorical variables into dummies and scale numerical predictor variables to a 0–1 (use function preprocess() with method = “range”—see Chapter 7). Generate a decile-wise lift chart for the training and validation sets. Interpret the meaning (in business terms) of the leftmost bar of the validation decile- wise lift chart.

b. Comment on the difference between the training and validation lift charts.

c. Run a second neural net model on the data, this time setting the number of hidden nodes to 1. Comment now on the difference between this model and the model you ran earlier, and how overfitting might have affected results.

d. What sort of information, if any, is provided about the effects of the various vari- ables?

CHAPTER 12

Discriminant Analysis

In this chapter, we describe the method of discriminant analysis, which is a model-based approach to classification. We discuss the main principle, where classification is based on the distance of a record from each of the class means. We explain the underlying measure of “statistical distance,” which takes into account the correlation between predictors. The output of a discriminant analysis proce- dure generates estimated “classification functions,” which are then used to pro- duce classification scores that can be translated into classifications or propensities (probabilities of class membership). One can also directly integrate misclassi- fication costs into the discriminant analysis setup, and we explain how this is achieved. Finally, we discuss the underlying model assumptions, the practical robustness to some assumption violations, and the advantages of discriminant analysis when the assumptions are reasonably met (for example, the sufficiency of a small training sample).

12.1 Introduction

Discriminant analysis is a classification method. Like logistic regression, it is a classical statistical technique that can be used for classification and profiling. It uses sets of measurements on different classes of records to classify new records into one of those classes (classification). Common uses of the method have been in classifying organisms into species and subspecies; classifying applications for loans, credit cards, and insurance into low- and high-risk categories; classifying customers of new products into early adopters, early majority, late majority, and laggards; classifying bonds into bond rating categories; classifying skulls of human fossils; as well as in research studies involving disputed authorship, decisions on college admission, medical studies involving alcoholics and nonalcoholics, and

293

294 DISCRIMINANT ANALYSIS

methods to identify human fingerprints. Discriminant analysis can also be used to highlight aspects that distinguish the classes (profiling).

We return to two examples that were described in earlier chapters, the Rid- ing Mowers and Personal Loan Acceptance examples. In each of these, the outcome variable has two classes. We close with a third example involving more than two classes.

Example 1: Riding Mowers

We return to the example from Chapter 7, where a riding mower manufacturer would like to find a way of classifying families in a city into those likely to purchase a riding mower and those not likely to purchase one. A pilot random sample of 12 owners and 12 nonowners in the city is undertaken. The data are given in Chapter 7 (Table 7.1), and a scatter plot is shown in Figure 12.1. We can think of a linear classification rule as a line that separates the two-dimensional region into two parts, with most of the owners in one half-plane and most nonowners in the complementary half-plane. A good classification rule would separate the data so that the fewest points are misclassified: The line shown in Figure 12.1 seems to do a good job in discriminating between the two classes as it makes four misclassifications out of 24 points. Can we do better?

FIGURE 12.1 SCATTER PLOT OF LOT SIZE VS. INCOME FOR 24 OWNERS AND NONOWNERS OF RIDING MOWERS. THE (AD HOC) LINE TRIES TO SEPARATE OWNERS FROM NONOWNERS

Example 2: Personal Loan Acceptance

The riding mowers example is a classic example and is useful in describing the concept and goal of discriminant analysis. However, in today’s business applica- tions, the number of records is much larger, and their separation into classes is much less distinct. To illustrate this, we return to the Universal Bank example

INTRODUCTION 295

FIGURE 12.2 PERSONAL LOAN ACCEPTANCE AS A FUNCTION OF INCOME AND CREDIT CARD SPENDING FOR 5000 CUSTOMERS OF THE UNIVERSAL BANK (IN LOG SCALE)

described in Chapter 9, where the bank’s goal is to identify new customers most likely to accept a personal loan. For simplicity, we will consider only two predic- tor variables: the customer’s annual income (Income, in $000s), and the average monthly credit card spending (CCAvg, in $000s). The first part of Figure 12.2 shows the acceptance of a personal loan by a subset of 200 customers from the bank’s database as a function of Income and CCAvg. We use a logarithmic scale on both axes to enhance visibility because there are many points condensed in the low-income, low-CC spending area. Even for this small subset, the sepa- ration is not clear. The second figure shows all 5000 customers and the added complexity of dealing with large numbers of records.

296 DISCRIMINANT ANALYSIS

12.2 Distance of a Record from a Class

Finding the best separation between records involves measuring their distance from their class. The general idea is to classify a record to the class to which it is closest. Suppose that we are required to classify a new customer of Universal Bank as being an acceptor or a nonacceptor of their personal loan offer, based on an income of x. From the bank’s database we find that the mean income for loan acceptors was $144.75K and for nonacceptors $66.24K. We can use Income as a predictor of loan acceptance via a simple Euclidean distance rule: If x is closer to the mean income of the acceptor class than to the mean income of the nonacceptor class, classify the customer as an acceptor; otherwise, classify the customer as a nonacceptor. In other words, if |x− 144.75| < |x− 66.24|, then classification = acceptor; otherwise, nonacceptor. Moving from a single predictor variable (income) to two or more predictor variables, the equivalent of the mean of a class is the centroid of a class. This is simply the vector of means x = [x1, . . . , xp]. The Euclidean distance between a record with p measurements x = [x1, . . . , xp] and the centroid x is defined as the square root of the sum of the squared differences between the individual values and the means:

DEuclidean(x, x) = √ (x1 − x1)2 + · · ·+ (xp − xp)2. (12.1)

Using the Euclidean distance has three drawbacks. First, the distance depends on the units we choose to measure the predictor variables. We will get different answers if we decide to measure income in dollars, for instance, rather than in thousands of dollars.

Second, Euclidean distance does not take into account the variability of the variables. For example, if we compare the variability in income in the two classes, we find that for acceptors, the standard deviation is lower than for nonacceptors ($31.6K vs. $40.6K). Therefore, the income of a new customer might be closer to the acceptors’ mean income in dollars, but because of the large variability in income for nonacceptors, this customer is just as likely to be a nonacceptor. We therefore want the distance measure to take into account the variance of the different variables and measure a distance in standard deviations rather than in the original units. This is equivalent to z-scores.

Third, Euclidean distance ignores the correlation between the variables. This is often a very important consideration, especially when we are using many pre- dictor variables to separate classes. In this case, there will often be variables, which by themselves are useful discriminators between classes, but in the pres- ence of other predictor variables are practically redundant, as they capture the same effects as the other variables.

A solution to these drawbacks is to use a measure called statistical distance (or Mahalanobis distance). Let us denote by S the covariance matrix between the p

FISHER’S LINEAR CLASSIFICATION FUNCTIONS 297

variables. The definition of a statistical distance is

DStatistical(x, x) = [x − x]′S−1[x − x]

= [(x1 − x1), (x2 − x2), . . . , (xp − xp)]S−1

x1 − x1 x2 − x2

... xp − xp

(12.2)

(the notation ′, which represents transpose operation, simply turns the column vector into a row vector). S−1 is the inverse matrix of S, which is the p- dimension extension to division. When there is a single predictor (p = 1), this formula reduces to a (squared) z-score calcualtion, since we subtract the mean and divide by the standard deviation. The statistical distance takes into account not only the predictor means, but also the spread of the predictor values and the correlations between the different predictors. To compute a statistical distance between a record and a class, we must compute the predictor means (the centroid) and the covariances between each pair of predictors. These are used to construct the distances. The method of discriminant analysis uses statistical distance as the basis for finding a separating line (or, if there are more than two variables, a separating hyperplane) that is equally distant from the different class means.1 It is based on measuring the statistical distances of a record to each of the classes and allocating it to the closest class. This is done through classification functions, which are explained next.

12.3 Fisher’s Linear Classification Functions

Linear classification functions were proposed in 1936 by the noted statistician R. A. Fisher as the basis for improved separation of records into classes. The idea is to find linear functions of the measurements that maximize the ratio of between-class variability to within-class variability. In other words, we would obtain classes that are very homogeneous and differ the most from each other. For each record, these functions are used to compute scores that measure the proximity of that record to each of the classes. A record is classified as belonging to the class for which it has the highest classification score (equivalent to the smallest statistical distance).

The classification functions are estimated using software. For example, Table 12.1 shows the classification functions obtained from running discriminant anal- ysis on the riding mowers data, using two predictors. Note that the number

1An alternative approach finds a separating line or hyperplane that is “best” at separating the different clouds of points. In the case of two classes, the two methods coincide.

298 DISCRIMINANT ANALYSIS

TABLE 12.1 DISCRIMINANT ANALYSIS FOR RIDING-MOWER DATA, DISPLAYING THE ESTIMATED CLASSIFICATION FUNCTIONS

code for linear discriminant analysis

library(DiscriMiner) mowers.df <- read.csv("RidingMowers.csv") da.reg <- linDA(mowers.df[,1:2], mowers.df[,3]) da.reg$functions

Output

> da.reg$functions Nonowner Owner

constant -51.4214499777 -73.1602116488 Income 0.3293554091 0.4295857129 Lot_Size 4.6815655074 5.4667502174

of classification functions is equal to the number of classes (in this case, two: owner/nonowner).

U S I N G C L A S S I F I C A T I O N F U N C T I O N S C O R E S T O C L A S S I F Y R E C O R D S

For each record, we calculate the value of the classification function (one for each class); whichever class’s function has the highest value (= score) is the class assigned to that record.

To classify a family into the class of owners or nonowners, we use the clas- sification functions to compute the family’s classification scores: A family is classified into the class of owners if the owner function score is higher than the nonowner function score, and into nonowners if the reverse is the case. These functions are specified in a way that can be easily generalized to more than two classes. The values given for the functions are simply the weights to be associated with each variable in the linear function in a manner anal- ogous to multiple linear regression. For instance, the first household has an income of $60K and a lot size of 18.4K ft2. Their owner score is therefore −73.16 + (0.43)(60) + (5.47)(18.4) = 53.2, and their nonowner score is −51.42 + (0.33)(60) + (4.68)(18.4) = 54.48. Since the second score is higher, the household is (mis)classified by the model as a nonowner. Such cal- culations are done by the software and do not need to be done manually. For example, the scores and classifications for all 24 households produced by R’s linDA() function are given in Table 12.2.

FISHER’S LINEAR CLASSIFICATION FUNCTIONS 299

TABLE 12.2 CLASSIFICATION SCORES, PREDICTED CLASSES, AND PROBABILITIES FOR RIDING-MOWER DATA

code for obtaining classification scores, predicted classes, and probabilities

da.reg <- linDA(mowers.df[,1:2], mowers.df[,3]) # compute probabilities manually (below); or, use lda() in package MASS with predict() propensity.owner <- exp(da.reg$scores[,2])/(exp(da.reg$scores[,1])+exp(da.reg$scores[,2])) data.frame(Actual=mowers.df$Ownership,

da.reg$classification, da.reg$scores, propensity.owner=propensity.owner)

Output

Actual da.reg.classification Nonowner Owner propensity.owner 1 Owner Nonowner 54.48067990 53.20313512 0.21796844638 2 Owner Owner 55.38873802 55.41077045 0.50550788510 3 Owner Owner 71.04259549 72.75874724 0.84763249334 4 Owner Owner 66.21047023 66.96771421 0.68075507343 5 Owner Owner 87.71741659 93.22905050 0.99597675014 6 Owner Owner 74.72663830 79.09877951 0.98753320250 7 Owner Owner 66.54448713 69.44984917 0.94811086581 8 Owner Owner 80.71624526 84.86469025 0.98445646665 9 Owner Owner 64.93538340 65.81620689 0.70699283969 10 Owner Owner 76.58516562 80.49966417 0.98043968864 11 Owner Owner 68.37011705 69.01716449 0.65634480272 12 Owner Owner 68.88764830 70.97123544 0.88929767077 13 Nonowner Owner 65.03888965 66.20702108 0.76280709688 14 Nonowner Nonowner 63.34507817 63.23031851 0.47134152980 15 Nonowner Nonowner 50.44370726 48.70504628 0.14948309577 16 Nonowner Nonowner 58.31064004 56.91959558 0.19924106677 17 Nonowner Owner 58.63995731 59.13979206 0.62242049465 18 Nonowner Nonowner 47.17838908 44.19020925 0.04796273478 19 Nonowner Nonowner 43.04730944 39.82518317 0.03834151004 20 Nonowner Nonowner 56.45681236 55.78064940 0.33711822839 21 Nonowner Nonowner 40.96767073 36.85685471 0.01612995010 22 Nonowner Nonowner 47.46071006 43.79102096 0.02485107730 23 Nonowner Nonowner 30.91759299 25.28316275 0.00355999342 24 Nonowner Nonowner 38.61511030 34.81159148 0.02180608595

An alternative way for classifying a record into one of the classes is to com- pute the probability of belonging to each of the classes and assigning the record to the most likely class. If we have two classes, we need only compute a single probability for each record (of belonging to owners, for example). Using a cutoff of 0.5 is equivalent to assigning the record to the class with the highest classifica- tion score. The advantage of this approach is that we obtain propensities, which can be used for goals such as ranking: we sort the records in order of descending probabilities and generate lift curves.

Let us assume that there arem classes. To compute the probability of belong- ing to a certain class k, for a certain record i, we need to compute all the clas- sification scores c1(i), c2(i), . . . , cm(i) and combine them using the following

300 DISCRIMINANT ANALYSIS

formula:

P[record i(with measurements x1, x2, ..., xp) belongs to class k]

= eck(i)

ec1(i) + ec2(i) + · · ·+ ecm(i) .

These probabilities and their computation for the riding mower data are given in Table 12.2. In R, function linDA() in package DiscriMiner provides scores and classifications, but does not directly provide propensities, so we need to use the above formula to compute probabilities from the scores (see R code in Table 12.2). Alternatively, use function lda in package MASS to automatically compute probabilities.

We now have three misclassifications, compared to four in our original (ad hoc) classification. This can be seen in Figure 12.3, which includes the line resulting from the discriminant model.2

FIGURE 12.3 CLASS SEPARATION OBTAINED FROM THE DISCRIMINANT MODEL (COMPARED TO AD HOC LINE FROM FIGURE 12.1)

12.4 Classification Performance of Discriminant Analysis

The discriminant analysis method relies on two main assumptions to arrive at classification scores: First, it assumes that the predictor measurements in all classes come from a multivariate normal distribution. When this assumption is reason- ably met, discriminant analysis is a more powerful tool than other classification

2The slope of the line is given by −a1/a2 and the intercept is a1/a2 x1 + x2, where ai is the difference between the ith classification function coefficients of owners and nonowners (e.g., here aincome = 0.43− 0.33).

CLASSIFICATION PERFORMANCE OF DISCRIMINANT ANALYSIS 301

methods, such as logistic regression. In fact, Efron (1975) showed that discrim- inant analysis is 30% more efficient than logistic regression if the data are mul- tivariate normal, in the sense that we require 30% less records to arrive at the same results. In practice, it has been shown that this method is relatively robust to departures from normality in the sense that predictors can be non-normal and even dummy variables. This is true as long as the smallest class is sufficiently large (approximately more than 20 records). This method is also known to be sensitive to outliers in both the univariate space of single predictors and in the multivari- ate space. Exploratory analysis should therefore be used to locate extreme cases and determine whether they can be eliminated.

The second assumption behind discriminant analysis is that the correlation structure between the different predictors within a class is the same across classes. This can be roughly checked by computing the correlation matrix between the predictors separately for each class and comparing matrices. If the correlations differ substantially across classes, the classifier will tend to classify records into the class with the largest variability. When the correlation structure differs signifi- cantly and the dataset is very large, an alternative is to use quadratic discriminant analysis.3

Notwithstanding the caveats embodied in these statistical assumptions, recall that in a predictive modeling environment, the ultimate test is whether the model works effectively. A reasonable approach is to conduct some exploratory anal- ysis with respect to normality and correlation, train and evaluate a model, then, depending on classification accuracy and what you learned from the initial explo- ration, circle back and explore further whether outliers should be examined or choice of predictor variables revisited.

With respect to the evaluation of classification accuracy, we once again use the general measures of performance that were described in Chapter 5 (judging the performance of a classifier), with the principal ones based on the confusion matrix (accuracy alone or combined with costs) for classification and the lift chart for ranking. The same argument for using the validation set for evaluating per- formance still holds. For example, in the riding mowers example, families 1, 13, and 17 are misclassified. This means that the model yields an error rate of 12.5% for these data. However, this rate is a biased estimate—it is overly optimistic, because we have used the same data for fitting the classification functions and for estimating the error. Therefore, as with all other models, we test performance on a validation set that includes data that were not involved in estimating the classification functions.

3In practice, quadratic discriminant analysis has not been found useful except when the difference in the correlation matrices is large and the number of records available for training and testing is large. The reason is that the quadratic model requires estimating many more parameters that are all subject to error [for m classes and p variables, the total number of parameters to be estimated for all the different correlation matrices is mp(p+ 1)/2].

302 DISCRIMINANT ANALYSIS

To obtain the confusion matrix from a discriminant analysis, we either use the classification scores directly or the propensities (probabilities of class member- ship) that are computed from the classification scores. In both cases, we decide on the class assignment of each record based on the highest score or probability. We then compare these classifications to the actual class memberships of these records. This yields the confusion matrix.

12.5 Prior Probabilities

So far we have assumed that our objective is to minimize the classification error. The method presented above assumes that the chances of encountering a record from either class is the same. If the probability of encountering a record for clas- sification in the future is not equal for the different classes, we should modify our functions to reduce our expected (long-run average) error rate. The modi- fication is done as follows: Let us denote by pj the prior or future probability of membership in class j (in the two-class case we have p1 and p2 = 1− p1). We modify the classification function for each class by adding log(pj). To illustrate this, suppose that the percentage of riding mower owners in the population is 15%, compared to 50% in the sample. This means that the model should classify fewer households as owners. To account for this distortion, we adjust the constants in the classification functions from Table 12.1 and obtain the adjusted constants −73.16+ log(0.15) = −75.06 for owners and −51.42+ log(0.85) = −50.58 for nonowners. To see how this can affect classifications, consider family 13, which was misclassified as an owner in the case involving equal probability of class membership. When we account for the lower probability of owning a mower in the population, family 13 is classified properly as a nonowner (its owner classification score is below the nonowner score).

12.6 Unequal Misclassification Costs

A second practical modification is needed when misclassification costs are not symmetrical. If the cost of misclassifying a class 1 record is very different from the cost of misclassifying a class 2 record, we may want to minimize the expected cost of misclassification rather than the simple error rate (which does not account for unequal misclassification costs). In the two-class case, it is easy to manipu- late the classification functions to account for differing misclassification costs (in addition to prior probabilities). We denote by q1 the cost of misclassifying a class 1 member (into class 2). Similarly, q2 denotes the cost of misclassifying a class 2 member (into class 1). These costs are integrated into the constants of the classi- fication functions by adding log(q1) to the constant for class 1 and log(q2) to the constant of class 2. To incorporate both prior probabilities and misclassification costs, add log(p1q1) to the constant of class 1 and log(p2q2) to that of class 2.

CLASSIFYING MORE THAN TWO CLASSES 303

In practice, it is not always simple to come up with misclassification costs q1 and q2 for each class. It is usually much easier to estimate the ratio of costs q2/q1 (e.g., the cost of misclassifying a credit defaulter is 10 times more expensive than that of misclassifying a nondefaulter). Luckily, the relationship between the classification functions depends only on this ratio. Therefore, we can set q1 = 1 and q2 = ratio and simply add log(q2/q1) to the constant for class 2.

12.7 Classifying More Than Two Classes

Example 3: Medical Dispatch to Accident Scenes

Ideally, every automobile accident call to the emergency number 911 results in the immediate dispatch of an ambulance to the accident scene. However, in some cases the dispatch might be delayed (e.g., at peak accident hours or in some resource-strapped towns or shifts). In such cases, the 911 dispatchers must make decisions about which units to send based on sketchy information. It is useful to augment the limited information provided in the initial call with additional information in order to classify the accident as minor injury, serious injury, or death. For this purpose, we can use data that were collected on automobile accidents in the United States in 2001 that involved some type of injury. For each accident, additional information is recorded, such as day of week, weather conditions, and road type. Figure 12.4 shows a small sample of records with 11 measurements of interest.

Accident # RushH

our WRK_ ZONE

WKDY INT_HWY LGTCON LEVEL SPD_

LIM SUR_C

OND TRAF_WAY WEATHER MAX_SEV

11 10 1 dark_light 701 ice one_way adverse no-injury

12 10 0 dark_light 700 ice divided adverse no-injury

13 10 0 dark_light 650 ice divided adverse non-fatal

14 10 0 dark_light 550 ice two_way not_adverse non-fatal

15 00 0 dark_light 350 snow one_way adverse no-injury

16 10 0 dark_light 351 wet divided adverse no-injury

07 10 1 dark_light 701 wet divided adverse non-fatal

08 10 0 dark_light 351 wet two_way adverse no-injury

19 10 0 dark_light 250 wet one_way adverse non-fatal

10 01 01 dark_light 350 wet divided adverse non-fatal

11 01 01 dark_light 300 wet divided adverse non-fatal

12 01 01 dark_light 600 wet divided not_adverse no-injury

13 01 01 dark_light 400 wet two_way not_adverse no-injury

14 00 01 day 651 dry two_way not_adverse fatal

15 01 00 day 550 dry two_way not_adverse fatal

16 01 01 day 550 dry two_way not_adverse non-fatal

17 01 00 day 550 dry two_way not_adverse non-fatal

18 00 01 dark 550 ice two_way not_adverse no-injury

19 00 00 dark 500 ice two_way adverse no-injury

20 00 00 dark 551 snow divided adverse no-injury

FIGURE 12.4 SAMPLE OF 20 AUTOMOBILE ACCIDENTS FROM THE 2001 DEPARTMENT OF TRANSPORTATION DATABASE. EACH ACCIDENT IS CLASSIFIED AS ONE OF THREE INJURY TYPES (NO-INJURY, NONFATAL, OR FATAL), AND HAS 10 MORE MEASUREMENTS (EXTRACTED FROM A LARGER SET OF MEASUREMENTS)

304 DISCRIMINANT ANALYSIS

The goal is to see how well the predictors can be used to classify injury type correctly. To evaluate this, a sample of 1000 records was drawn and partitioned into training and validation sets, and a discriminant analysis was performed on the training data. The output structure is very similar to that for the two-class case. The only difference is that each record now has three classification functions (one for each injury type), and the confusion and error matrices are of size 3×3 to account for all the combinations of correct and incorrect classifications (see Table 12.3). The rule for classification is still to classify a record to the class that has the highest corresponding classification score. The classification scores are computed, as before, using the classification function coefficients. This can

TABLE 12.3 DISCRIMINANT ANALYSIS FOR THE THREE-CLASS INJURY EXAMPLE: CLASSIFICATION FUNCTIONS AND CONFUSION MATRIX FOR TRAINING SET

code for running linear discriminant analysis on the accidents data

library(DiscriMiner) library(caret)

accidents.df <- read.csv("Accidents.csv") da.reg <- linDA(accidents.df[,1:10], accidents.df[,11]) da.reg$functions confusionMatrix(da.reg$classification, accidents.df$MAX_SEV)

Output

> da.reg$functions fatal no-injury non-fatal

constant -25.5958095610 -24.514323034 -24.2336221574 RushHour 0.9225623509 1.952403425 1.9031991672 WRK_ZONE 0.5178609440 1.195060274 0.7705682214 WKDY 4.7801494470 6.417633787 6.1165223679 INT_HWY -1.8418782656 -2.673037935 -2.5366224810 LGTCON_day 3.7070124215 3.666075602 3.7276207831 LEVEL 2.6268937732 1.567550702 1.7138656960 SPD_LIM 0.5051317221 0.461479676 0.4520847732 SUR_COND_dry 9.9988600752 15.833794528 16.2565639740 TRAF_two_way 7.1079766143 6.342147286 6.3549435330 WEATHER_adverse 9.6880211017 16.363876853 16.3172755675

> confusionMatrix(da.reg$classification, accidents.df$MAX_SEV) Confusion Matrix and Statistics

Reference Prediction fatal no-injury non-fatal

fatal 1 6 6 no-injury 1 114 95 non-fatal 3 172 202

Overall Statistics

Accuracy : 0.5283333

CLASSIFYING MORE THAN TWO CLASSES 305

TABLE 12.4 CLASSIFICATION SCORES, MEMBERSHIP PROBABILITIES, AND CLASSIFICATIONS FOR THE THREE-CLASS INJURY TRAINING DATASET

code for producing linear discriminant analysis scores and propensities

prob <- exp(da.reg$scores[,1:3])/ (exp(da.reg$scores[,1])+exp(da.reg$scores[,2])+exp(da.reg$scores[,3]))

res <- data.frame(Classification = lda.reg$classification, Actual = accidents.df$MAX_SEV, Score = round(da.reg$scores,2), Propensity = round(propensity,2))

head(res)

Output

> head(res) Classification Actual Score.fatal Score.no.injury Score.non.fatal

1 no-injury no-injury 25.94 31.42 30.93 2 no-injury non-fatal 15.00 15.58 15.01 3 no-injury no-injury 2.69 9.95 9.81 4 no-injury no-injury 10.10 17.94 17.64 5 no-injury non-fatal 2.42 11.76 11.41 6 no-injury non-fatal 7.47 16.37 15.93 Propensity.fatal Propensity.no.injury Propensity.non.fatal

1 0.00 0.62 0.38 2 0.26 0.47 0.27 3 0.00 0.54 0.46 4 0.00 0.57 0.43 5 0.00 0.59 0.41 6 0.00 0.61 0.39

be seen in Table 12.4. For instance, the no-injury classification score for the first accident in the training set is −24.51 + (1.95)(1) + (1.19)(0) + · · · + (16.36)(1) = 31.42. The nonfatal score is similarly computed as 30.93 and the fatal score as 25.94. Since the no-injury score is highest, this accident is (correctly) classified as having no injuries.

We can also compute for each accident, the propensities (estimated probabil- ities) of belonging to each of the three classes using the same relationship between classification scores and probabilities as in the two-class case. For instance, the probability of the above accident involving nonfatal injuries is estimated by the model as

e30.93

e31.42 + e30.93 + e25.94 = 0.38. (12.3)

The probabilities of an accident involving no injuries or fatal injuries are com- puted in a similar manner. For the first accident in the training set, the highest probability is that of involving no injuries, and therefore it is classified as a no- injury accident.

306 DISCRIMINANT ANALYSIS

12.8 Advantages and Weaknesses

Discriminant analysis is typically considered more of a statistical classification method than a data mining method. This is reflected in its absence or short mention in many data mining resources. However, it is very popular in social sciences and has shown good performance. The use and performance of dis- criminant analysis are similar to those of multiple linear regression. The two methods therefore share several advantages and weaknesses.

Like linear regression, discriminant analysis searches for the optimal weight- ing of predictors. In linear regression, weighting is with relation to the numerical outcome variable, whereas in discriminant analysis, it is with relation to separat- ing the classes. Both use least squares for estimation and the resulting estimates are robust to local optima.

In both methods, an underlying assumption is normality. In discriminant analysis, we assume that the predictors are approximately from a multivariate normal distribution. Although this assumption is violated in many practical sit- uations (such as with commonly-used binary predictors), the method is surpris- ingly robust. According to Hastie et al. (2001), the reason might be that data can usually support only simple separation boundaries, such as linear boundaries. However, for continuous variables that are found to be very skewed (as can be seen through a histogram), transformations such as the log transform can improve performance. In addition, the method’s sensitivity to outliers commands explor- ing the data for extreme values and removing those records from the analysis.

An advantage of discriminant analysis as a classifier (like logistic regression in this respect) is that it provides estimates of single-predictor contributions.4

This is useful for obtaining a ranking of predictor importance, and for variable selection.

Finally, the method is computationally simple, parsimonious, and especially useful for small datasets. With its parametric form, discriminant analysis makes the most out of the data and is therefore especially useful with small samples (as explained in Section 12.4).

4Comparing predictor contribution requires normalizing all the predictors before running discriminant analysis. Then, compare each coefficient across the two classification functions: coefficients with large differences indicate a predictor with high separation power.

PROBLEMS 307

PROBLEMS

12.1 Personal Loan Acceptance. Universal Bank is a relatively young bank growing rapidly in terms of overall customer acquisition. The majority of these customers are liability customers with varying sizes of relationship with the bank. The customer base of asset customers is quite small, and the bank is interested in expanding this base rapidly to bring in more loan business. In particular, it wants to explore ways of converting its liability customers to personal loan customers.

A campaign the bank ran for liability customers last year showed a healthy conver- sion rate of over 9% successes. This has encouraged the retail marketing department to devise smarter campaigns with better target marketing. The goal of our analysis is to model the previous campaign’s customer behavior to analyze what combination of factors make a customer more likely to accept a personal loan. This will serve as the basis for the design of a new campaign.

The file UniversalBank.csv contains data on 5000 customers. The data include customer demographic information (e.g., age, income), the customer’s relationship with the bank (e.g., mortgage, securities account), and the customer response to the last personal loan campaign (Personal Loan). Among these 5000 customers, only 480 (= 9.6%) accepted the personal loan that was offered to them in the previous campaign.

Partition the data (60% training and 40% validation) and then perform a discrim- inant analysis that models Personal Loan as a function of the remaining predictors (excluding zip code). Remember to turn categorical predictors with more than two categories into dummy variables first. Specify the success class as 1 (personal loan acceptance), and use the default cutoff value of 0.5.

a. Compute summary statistics for the predictors separately for loan acceptors and nonacceptors. For continuous predictors, compute the mean and standard devi- ation. For categorical predictors, compute the percentages. Are there predictors where the two classes differ substantially?

b. Examine the model performance on the validation set.

i. What is the accuracy rate?

ii. Is one type of misclassification more likely than the other?

iii. Select three customers who were misclassified as acceptors and three who were misclassified as nonacceptors. The goal is to determine why they are misclassified. First, examine their probability of being classified as acceptors: is it close to the threshold of 0.5? If not, compare their predictor values to the summary statistics of the two classes to determine why they were misclassified.

c. As in many marketing campaigns, it is more important to identify customers who will accept the offer rather than customers who will not accept it. Therefore, a good model should be especially accurate at detecting acceptors. Examine the lift chart and decile-wise lift chart for the validation set and interpret them in light of this ranking goal.

d. Compare the results from the discriminant analysis with those from a logistic regres- sion (both with cutoff 0.5 and the same predictors). Examine the confusion matri- ces, the lift charts, and the decile charts. Which method performs better on your validation set in detecting the acceptors?

e. The bank is planning to continue its campaign by sending its offer to 1000 additional customers. Suppose that the cost of sending the offer is $1 and the profit from an accepted offer is $50. What is the expected profitability of this campaign?

308 DISCRIMINANT ANALYSIS

f. The cost of misclassifying a loan acceptor customer as a nonacceptor is much higher than the opposite misclassification cost. To minimize the expected cost of misclas- sification, should the cutoff value for classification (which is currently at 0.5) be increased or decreased?

12.2 Identifying Good System Administrators. A management consultant is study- ing the roles played by experience and training in a system administrator’s ability to complete a set of tasks in a specified amount of time. In particular, she is interested in discriminating between administrators who are able to complete given tasks within a specified time and those who are not. Data are collected on the performance of 75 randomly selected administrators. They are stored in the file SystemAdministrators.csv.

Using these data, the consultant performs a discriminant analysis. The vari- able Experience measures months of full time system administrator experience, while Training measures number of relevant training credits. The dependent variable Com- pleted is either Yes or No, according to whether or not the administrator completed the tasks.

a. Create a scatter plot of Experience vs. Training using color or symbol to differ- entiate administrators who completed the tasks from those who did not complete them. See if you can identify a line that separates the two classes with minimum misclassification.

b. Run a discriminant analysis with both predictors using the entire dataset as training data. Among those who completed the tasks, what is the percentage of adminis- trators who are classified incorrectly as failing to complete the tasks?

c. Compute the two classification scores for an administrator with 4 months of expe- rience and 6 credits of training. Based on these, how would you classify this admin- istrator?

d. How much experience must be accumulated by an administrator with 4 training credits before his or her estimated probability of completing the tasks exceeds 0.5?

e. Compare the classification accuracy of this model to that resulting from a logistic regression with cutoff 0.5.

12.3 Detecting Spam E-mail (from the UCI Machine Learning Repository). A team at Hewlett-Packard collected data on a large number of e-mail messages from their postmaster and personal e-mail for the purpose of finding a classifier that can separate e-mail messages that are spam vs. nonspam (a.k.a. “ham”). The spam con- cept is diverse: It includes advertisements for products or websites, “make money fast” schemes, chain letters, pornography, and so on. The definition used here is “unso- licited commercial e-mail.” The file Spambase.csv contains information on 4601 e-mail messages, among which 1813 are tagged “spam.” The predictors include 57 attributes, most of them are the average number of times a certain word (e.g., mail, George) or symbol (e.g., #, !) appears in the e-mail. A few predictors are related to the number and length of capitalized words.

a. To reduce the number of predictors to a manageable size, examine how each pre- dictor differs between the spam and nonspam e-mails by comparing the spam-class average and nonspam-class average. Which are the 11 predictors that appear to vary the most between spam and nonspam e-mails? From these 11, which words or signs occur more often in spam?

b. Partition the data into training and validation sets, then perform a discriminant analysis on the training data using only the 11 predictors.

PROBLEMS 309

c. If we are interested mainly in detecting spam messages, is this model useful? Use the confusion matrix, lift chart, and decile chart for the validation set for the evaluation.

d. In the sample, almost 40% of the e-mail messages were tagged as spam. However, suppose that the actual proportion of spam messages in these e-mail accounts is 10%. Compute the constants of the classification functions to account for this information.

e. A spam filter that is based on your model is used, so that only messages that are classified as nonspam are delivered, while messages that are classified as spam are quarantined. In this case, misclassifying a nonspam e-mail (as spam) has much heftier results. Suppose that the cost of quarantining a nonspam e-mail is 20 times that of not detecting a spam message. Compute the constants of the classification functions to account for these costs (assume that the proportion of spam is reflected correctly by the sample proportion).

CHAPTER 13

Combining Methods: Ensembles and Uplift Modeling

In this chapter, we look at two useful approaches that combine methods for improving predictive power: ensembles and uplift modeling. An ensemble com- bines multiple supervised models into a “super-model.” The previous chapters in this part of the book introduced different supervised methods for prediction and classification. Earlier, in Chapter 5, we learned about evaluating predic- tive performance, which can be used to compare several models and choose the best one. An ensemble is based on the powerful notion of combining models. Instead of choosing a single predictive model, we can combine several models to achieve improved predictive accuracy. In this chapter, we explain the under- lying logic of why ensembles can improve predictive accuracy and introduce popular approaches for combining models, including simple averaging, bagging, and boosting.

In uplift modeling, we combine supervised modeling with A-B testing, which is a simple type of a randomized experiment. We describe the basics of A-B testing and how it is used along with predictive models in persuasion messaging not to predict outcomes but to predict who should receive which message or treatment.

13.1 Ensembles1

Ensembles played a major role in the million-dollar Netflix Prize contest that started in 2006. At the time, Netflix, the largest DVD rental service in the

1This and subsequent sections in this chapter copyright © 2017 Datastats, LLC, and Galit Shmueli. Used by permission.

311

312 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

United States, wanted to improve their movie recommendation system (from www.netflixprize.com):

Netflix is all about connecting people to the movies they love. To help customers find those movies, we’ve developed our world-class movie recommendation system: CinematchSM…And while Cinematch is doing pretty well, it can always be made better.

In a bold move, the company decided to share a large amount of data on movie ratings by their users, and set up a contest, open to the public, aimed at improving their recommendation system:

We provide you with a lot of anonymous rating data, and a prediction accuracy bar that is 10% better than what Cinematch can do on the same training data set.

During the contest, an active leader-board showed the results of the com- peting teams. An interesting behavior started appearing: Different teams joined forces to create combined, or ensemble predictions, which proved more accurate than the individual predictions. The winning team, called “BellKor’s Pragmatic Chaos” combined results from the “BellKor” and “Big Chaos” teams alongside additional members. In a 2010 article in Chance magazine, the Netflix Prize winners described the power of their ensemble approach:

An early lesson of the competition was the value of combining sets of predictions from multiple models or algorithms. If two prediction sets achieved similar RMSEs, it was quicker and more effective to simply average the two sets than to try to develop a new model that incorporated the best of each method. Even if the RMSE for one set was much worse than the other, there was almost certainly a linear combination that improved on the better set.

Why Ensembles Can Improve Predictive Power

The principle of combining methods is popular for reducing risk. For example, in finance, portfolios are created for reducing investment risk. The return from a portfolio is typically less risky, because the variation is smaller than each of the individual components.

In predictive modeling, “risk” is equivalent to variation in prediction error. The more our prediction errors vary, the more volatile our predictive model. Consider predictions from two different models for a set of n records. e1,i is the prediction error for the ith record by method 1 and e2,i is the prediction error for the same record by method 2.

Suppose that each model produces prediction errors that are, on average, zero (for some records the model over-predicts and for some it under-predicts,

ENSEMBLES 313

but on average the error is zero):

E(e1,i) = E(e2,i) = 0.

If, for each record, we take an average of the two predictions: yi = ŷ1,i+ŷ2,i

2 ,

then the expected mean error will also be zero:

E (yi − yi) = E ( yi −

ŷ1,i + ŷ2,i 2

) (13.1)

= E

( yi − ŷ1,i

2 +

yi − ŷ2,i 2

) = E

( e1,i + e2,i

2

) = 0.

This means that the ensemble has the same mean error as the individual models. Now let us examine the variance of the ensemble’s prediction errors:

Var

( e1,i + e2,i

2

) =

1

4 (Var(e1,i) + Var(e2,i)) +

1

4 × 2Cov (e1,i, e2,i) .

(13.2) This variance can be lower than each of the individual variances Var(e1,i) and Var(e2,i) under some circumstances. A key component is the covariance (or equivalently, correlation) between the two prediction errors. The case of no correlation leaves us with a quantity that can be smaller than each of the individ- ual variances. The variance of the average prediction error will be even smaller when the two prediction errors are negatively correlated.

In summary, using an average of two predictions can potentially lead to smaller error variance, and therefore better predictive power. These results gen- eralize to more than two methods; you can combine results from multiple pre- diction methods or classifiers.

T H E W I S D O M O F C R O W D S

In his book The Wisdom of Crowds, James Surowiecki recounts how Francis Galton, a prominent statistician from the 19th century, watched a contest at a county fair in England. The contest’s objective was to guess the weight of an ox. Individual contest entries were highly variable, but the mean of all the estimates was surpris- ingly accurate—within 1% of the true weight of the ox. On balance, the errors from multiple guesses tended to cancel one another out. You can think of the output of a predictive model as a more informed version of these guesses. Averaging together multiple guesses will yield a more precise answer than the vast majority of the indi- vidual guesses. Note that in Galton’s story, there were a few (lucky) individuals who scored better than the average. An ensemble estimate will not always be more accurate than all the individual estimates in all cases, but it will be more accurate most of the time.

314 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

Simple Averaging

The simplest approach for creating an ensemble is to combine the predictions, classifications, or propensities from multiple models. For example, we might have a linear regression model, a regression tree, and a k-NN algorithm. We use each of the three methods to score, say, a test set. We then combine the three sets of results.

The three models can also be variations that use the same algorithm. For example, we might have three linear regression models, each using a different set of predictors.

Combining Predictions In prediction tasks, where the outcome variable is numerical, we can combine the predictions from the different methods simply by taking an average. In the above example, for each record in the test set, we have three predictions (one from each model). The ensemble prediction is then the average of the three values.

One alternative to a simple average is taking the median prediction, which would be less affected by extreme predictions. Another possibility is computing a weighted average, where weights are proportional to a quantity of interest. For instance, weights can be proportional to the accuracy of the model, or if differ- ent data sources are used, the weights can be proportional to the quality of the data.

Ensembles for prediction are useful not only in cross-sectional prediction, but also in time series forecasting (see Chapters 16–18). In forecasting, the same approach of combining future forecasts from multiple methods can lead to more precise predictions. One example is the weather forecasting application Fore- cast.io (www.forecast.io), which describes their algorithm as follows:

Forecast.io is backed by a wide range of data sources, which are aggregated together statistically to provide the most accurate forecast possible for a given location.

Combining Classifications In the case of classification, combining the results from multiple classifiers can be done using “voting”: For each record, we have multiple classifications. A simple rule would be to choose the most popular class among these classifications. For example, we might use a classification tree, a naive Bayes classifier, and discriminant analysis for classifying a binary outcome. For each record we then generate three predicted classes. Simple voting would choose the most common class among the three.

As in prediction, we can assign heavier weights to scores from some models, based on considerations such as model accuracy or data quality. This would be done by setting a “majority rule” that is different from 50%.

ENSEMBLES 315

Combining Propensities Similar to predictions, propensities can be combined by taking a simple (or weighted) average. Recall that some algo- rithms, such as naive Bayes (see Chapter 8), produce biased propensities and should therefore not be simply averaged with propensities from other methods.

Bagging

Another form of ensembles is based on averaging across multiple random data samples. Bagging, short for “bootstrap aggregating,” comprises two steps:

1. Generate multiple random samples (by sampling with replacement from the original data)—this method is called “bootstrap sampling.”

2. Running an algorithm on each sample and producing scores.

Bagging improves the performance stability of a model and helps avoid over- fitting by separately modeling different data samples and then combining the results. It is therefore especially useful for algorithms such as trees and neural networks.

Boosting

Boosting is a slightly different approach to creating ensembles. Here the goal is to directly improve areas in the data where our model makes errors, by forcing the model to pay more attention to those records. The steps in boosting are:

1. Fit a model to the data.

2. Draw a sample from the data so that misclassified records (or records with large prediction errors) have higher probabilities of selection.

3. Fit the model to the new sample.

4. Repeat Steps 2–3 multiple times.

Bagging and Boosting in R

In Chapter 9, we described random forests, an ensemble based on bagged trees. We illustrated a random forest implementation for the personal loan example. The adabag package in R can be used to generate bagged and boosted trees. Tables 13.1 and 13.2 show the R code and output producing a bagged tree and a boosted tree for the personal loan data, and how they are used to generate classifications for the validation set.

Advantages and Weaknesses of Ensembles

Combining scores from multiple models is aimed at generating more precise pre- dictions (lowering the prediction error variance). The ensemble approach is most useful when the combined models generate prediction errors that are negatively

316 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

TABLE 13.1 EXAMPLE OF BAGGING AND BOOSTING CLASSIFICATION TREES ON THE PERSONAL LOAN DATA: R CODE

code for bagging and boosing trees

library(adabag) library(rpart) library(caret)

bank.df <- read.csv("UniversalBank.csv") bank.df <- bank.df[ , -c(1, 5)] # Drop ID and zip code columns.

# transform Personal.Loan into categorical variable bank.df$Personal.Loan = as.factor(bank.df$Personal.Loan)

# partition the data train.index <- sample(c(1:dim(bank.df)[1]), dim(bank.df)[1]*0.6) train.df <- bank.df[train.index, ] valid.df <- bank.df[-train.index, ]

# single tree tr <- rpart(Personal.Loan ~ ., data = train.df) pred <- predict(tr, valid.df, type = "class") confusionMatrix(pred, valid.df$Personal.Loan)

# bagging bag <- bagging(Personal.Loan ~ ., data = train.df) pred <- predict(bag, valid.df, type = "class") confusionMatrix(pred$class, valid.df$Personal.Loan)

# boosting boost <- boosting(Personal.Loan ~ ., data = bank.df) pred <- predict(boost, valid.df, type = "class") confusionMatrix(pred$class, valid.df$Personal.Loan)

associated, but it can also be useful when the correlation is low. Ensembles can use simple averaging, weighted averaging, voting, medians, etc. Models can be based on the same algorithm or on different algorithms, using the same sample or different samples. Ensembles have become a major strategy for participants in data mining contests, where the goal is to optimize some predictive measure. In that sense, ensembles also provide an operational way to obtain solutions with high predictive power in a fast way, by engaging multiple teams of “data crunch- ers” working in parallel and combining their results.

Ensembles that are based on different data samples help avoid overfitting. However, remember that you can also overfit the data with an ensemble if you tweak it (e.g., choosing the “best” weights when using a weighted average).

The major disadvantage of an ensemble is the resources that it requires: com- putationally, as well as in terms of software availability and the analyst’s skill and time investment. Ensembles that combine results from different algorithms

UPLIFT (PERSUASION) MODELING 317

TABLE 13.2 EXAMPLE OF BAGGING AND BOOSTING CLASSIFICATION TREES ON THE PERSONAL LOAN DATA: OUTPUT

> # single tree > confusionMatrix(pred, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1792 18 1 19 171

Accuracy : 0.9815

> # bagging > confusionMatrix(pred$class, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1797 23 1 14 166

Accuracy : 0.9815

> # boosting > confusionMatrix(pred$class, valid.df$Personal.Loan) Confusion Matrix and Statistics

Reference Prediction 0 1

0 1804 16 1 7 173

Accuracy : 0.9885

require developing each of the models and evaluating them. Boosting-type ensembles and bagging-type ensembles do not require such effort, but they do have a computational cost (although boosting can be parallelized easily). Ensem- bles that rely on multiple data sources require collecting and maintaining mul- tiple data sources. And finally, ensembles are “blackbox” methods, in that the relationship between the predictors and the outcome variable usually becomes nontransparent.

13.2 Uplift (Persuasion) Modeling

Long before the advent of the Internet, sending messages directly to individuals (i.e., direct mail) held a big share of the advertising market. Direct marketing affords the marketer the ability to invite and monitor direct responses from con- sumers. This, in turn, allows the marketer to learn whether the messaging is paying off. A message can be tested with a small section of a large list and, if it

318 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

pays off, the message can be rolled out to the entire list. With predictive model- ing, we have seen that the rollout can be targeted to that portion of the list that is most likely to respond or behave in a certain way. None of this was possible with traditional media advertising (television, radio, newspaper, magazine).

Direct response also made it possible to test one message against another and find out which does better.

A-B Testing

A-B testing is the marketing industry’s term for a standard scientific experiment in which results can be tracked for each individual. The idea is to test one treatment against another, or a treatment against a control. “Treatment” is simply the term for the intervention you are testing: In a medical trial it is typically a drug, device, or other therapy; in marketing it is typically an offering to a consumer—for example, an e-mail, or a web page shown to a consumer. A general display ad in a magazine would not generally qualify, unless it had a specific call to action that allowed the marketer to trace the action (e.g., purchase) to a given ad, plus the ability to split the magazine distribution randomly and provide a different offer to each segment.

An important element of A-B testing is random allocation—the treatments are assigned or delivered to individuals randomly. That way, any difference between treatment A and treatment B can be attributed to the treatment (unless it is due to chance).

Uplift

An A-B test tells you which treatment does better on average, but says nothing about which treatment does better for which individual. A classic example is in political campaigns. Consider the following scenario: The campaign director for Smith, a Democratic Congressional candidate, would like to know which voters should be called to encourage to support Smith. Voters that tend to vote Democratic but are not activists might be more inclined to vote for Smith if they got a call. Active Democrats are probably already supportive of him, and therefore a call to them would be wasted. Calls to Republicans are not only wasteful, but they could be harmful.

Campaigns now maintain extensive data on voters to help guide decisions about outreach to individual voters. Prior to the 2008 Obama campaign, the practice was to make rule-based decisions based on expert political judgment. Since 2008, it has increasingly been recognized that, rather than relying on judg- ment or supposition to determine whether an individual should be called, it is best to use the data to develop a model that can predict whether a voter will respond positively to outreach.

UPLIFT (PERSUASION) MODELING 319

Gathering the Data

US states maintain publicly available files of voters, as part of the transparent oversight process for elections. The voter file contains data such as name, address, and date of birth. Political parties have “poll-watchers” at elections to record who votes, so they have additional data on which elections voters voted in. Census data for neighborhoods can be appended, based on voter address. Finally, commercial demographic data can be purchased and matched to the voter data. Table 13.3 shows a small extract of data derived from the voter file for the US state of Delaware.2 The actual data used in this problem are in the file Persuasion-A.csv and contain 10,000 records and many additional variables beyond those shown in Table 13.3.

First, the campaign director conducts a survey of 10,000 voters to deter- mine their inclination to vote Democratic. Then she conducts an experiment, randomly splitting the sample of 10,000 voters in half and mailing a message promoting Smith to half the list (treatment A), and nothing to the other half (treatment B). The control group that gets no message is essential, since other campaigns or news events might cause a shift in opinion. The goal is to measure the change in opinion after the message is sent out, relative to the no-message control group.

The next step is conducting a post-message survey of the same sample of 10,000 voters, to measure whether each voter’s opinion of Smith has shifted in

TABLE 13.3 DATA ON VOTERS (SMALL SUBSET OF VARIABLES AND RECORDS) AND DATA DICTIONARY

Voter Age NH_White Comm_PT H_F1 Reg_Days PR_Pelig E_Elig Political_C

1 28 70 0 0 3997 0 20 1 2 23 67 3 0 300 0 0 1 3 57 64 4 0 2967 0 0 0 4 70 53 2 1 16620 100 90 1 5 37 76 2 0 3786 0 20 0

Data Dictionary

Age Voter age in years NH_White Neighborhood average of % non-Hispanic white in household Comm_PT Neighborhood % of workers who take public transit H_F1 Single female household (1 = yes) Reg_Days Days since voter registered at current address PR_Pelig Voted in what % of non-presidential primaries E_Pelig Voted in what % of any primaries Political_C Is there a political contributor in the home? (1 = yes)

2Thanks to Ken Strasma, founder of the microtargeting firm HaystaqDNA and director of targeting for the 2004 Kerry campaign and the 2008 Obama campaign, for these data.

320 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

a positive direction. A binary variable, Moved_AD, will be added to the above data, indicating whether opinion has moved in a Democratic direction (1) or not (0).

Table 13.4 summarizes the results of the survey, by comparing the move- ment in a Democratic direction for each of the treatments. Overall, the message (Message = 1) is modestly effective.

Movement in a Democratic direction among those who got no message is 34.4%. This probably reflects the approach of the election, the heightening cam- paign activity, and the reduction in the “no opinion” category. It also illustrates the need for a control group. Among those who did get the message, the move- ment in a Democratic direction is 40.2%. So, overall, the lift from the message is 5.8%.

TABLE 13.4 RESULTS OF SENDING A PRO-DEMOCRATIC MESSAGE TO VOTERS

#Voters # Moved Dem. % Moved Dem.

Message = 1 (message sent) 5000 2012 40.2% Message = 0 (no message sent) 5000 1722 34.4%

We can now append two variables to the voter data shown earlier in Table 13.3: message [whether they received the message (1) or not (0)] and Moved_AD [whether they moved in a Democratic direction (1) or not (0)]. The augmented data are shown in Table 13.5.

TABLE 13.5 OUTCOME VARIABLE (MOVED_AD) AND TREATMENT VARIABLE (MESSAGE) ADDED TO VOTER DATA

Voter Age NH_White Comm_PT H_F1 Reg_Days PR_Pelig E_Elig Political_C Message Moved_AD

1 28 70 0 0 3997 0 20 1 0 1 2 23 67 3 0 300 0 0 1 1 1 3 57 64 4 0 2967 0 0 0 0 0 4 70 53 2 1 16620 100 90 1 0 0 5 37 76 2 0 3786 0 20 0 1 0

A Simple Model

We can develop a predictive model with Moved_AD as the outcome variable, and various predictor variables, including the treatment Message. Any classification method can be used; Table 13.6 shows the first few lines from the output of a logistic regression model used to predict Moved_AD.

However, our interest is not just how the message did overall, nor is it whether we can predict the probability that a voter’s opinion will move in a favorable direction. Rather our goal is to predict how much (positive) impact the message will have on a specific voter. That way the campaign can direct

UPLIFT (PERSUASION) MODELING 321

TABLE 13.6 CLASSIFICATIONS AND PROPENSITIES FROM PREDICTIVE MODEL (SMALL EXTRACT)

Voter Message Actual Moved_AD Predicted Moved_AD Predicted Prob.

1 0 1 1 0.5975 2 1 1 1 0.5005 3 0 0 0 0.2235 4 0 0 0 0.3052 5 1 0 0 0.4140

its limited resources toward the voters who are the most persuadable—those for whom sending the message will have the greatest positive effect.

Modeling Individual Uplift

To answer the question about the message’s impact on each voter, we need to model the effect of the message at the individual voter level. For each voter, uplift is defined as follows:

Uplift = increase in propensity of favorable opinion after receiving message To build an uplift model, we follow the following steps to estimate the change

in probability of “success” (propensity) that comes from receiving the treatment (the message):

1. Randomly split a data sample into treatment and control groups, conduct an A-B test, and record the outcome (in our example: Moved_AD)

2. Recombining the data sample, partition it into training and validation sets; build a predictive model with this outcome variable and include a predictor variable that denotes treatment status (in our example: Message)

3. Score this predictive model to a partition of the data; you can use the vali- dation partition. This will yield, for each validation record, its propensity of success given its treatment.

4. Reverse the value of the treatment variable and re-score the same model to that partition. This will yield for each validation record its propensity of success had it received the other treatment.

5. Uplift is estimated for each individual by P(Success | Treatment = 1) − P(Success | Treatment = 0)

6. For new data where no experiment has been performed, simply add a synthetic predictor variable for treatment and assign first a ‘1,’ score the model, then a ‘0,’ and score the model again. Estimate uplift for the new record(s) as above.

Continuing with the small voter example, the results from Step 3 were shown in Table 13.6—the right column shows the propensities from the model. Next,

322 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

TABLE 13.7 CLASSIFICATIONS AND PROPENSITIES FROM PREDICTIVE MODEL (SMALL EXTRACT) WITH MESSAGE VALUES REVERSED

Voter Message Actual Moved_AD Predicted Moved_AD Predicted Prob.

1 1 1 1 0.6908 2 0 1 1 0.3996 3 1 0 0 0.3022 4 1 0 0 0.3980 5 0 0 0 0.3194

we re-estimate the logistic model, but with the values of the treatment variable Message reversed for each row. Table 13.7 shows the propensities with variable Message reversed (you can see the reversed values in column Message). Finally, in Step 5, we calculate the uplift for each voter.

Table 13.8 shows the uplift for each voter—the success (Moved_AD = 1) propensity given Message = 1 minus the success propensity given Message = 0.

TABLE 13.8 UPLIFT: CHANGE IN PROPENSITIES FROM SENDING MESSAGE VS. NOT SENDING MESSAGE

Voter Prob. if Message = 1 Prob. if Message = 0 Uplift

1 0.6908 0.5975 0.0933 2 0.5005 0.3996 0.1009 3 0.3022 0.2235 0.0787 4 0.3980 0.3052 0.0928 5 0.4140 0.3194 0.0946

Computing Uplift with R

This entire process that we showed manually can be done using R’s uplift package. One difference between our manual computation and R’s uplift pack- age is that we used a logistic regression whereas the uplift package uses either a random forest (upliftRF()) or a k-nearest neighbors classifier (upliftKNN()). Table 13.9 shows the result of implementing uplift analysis in R using a random forest. The output shows the two conditional probabilities, P(Success | Treatment = 1) and P(Success | Treatment = 0), estimated for each record. The difference between the values in Tables 13.8 and 13.9 are due to the use of two differ- ent predictive algorithms (logistic regression vs. random forests).

Using the Results of an Uplift Model

Once we have estimated the uplift for each individual, the results can be ordered by uplift. The message could then be sent to all those voters with a positive uplift, or, if resources are limited, only to a subset—those with the greatest uplift.

UPLIFT (PERSUASION) MODELING 323

TABLE 13.9 UPLIFT IN R APPLIED TO THE VOTERS DATA

code for uplift

library(uplift) voter.df <- read.csv("Voter-Persuasion.csv") # transform variable MOVED_AD to numerical voter.df$MOVED_AD_NUM <- ifelse(voter.df$MOVED_AD == "Y", 1, 0)

set.seed(1) train.index <- sample(c(1:dim(voter.df)[1]), dim(voter.df)[1]*0.6) train.df <- voter.df[train.index, ] valid.df <- voter.df[-train.index, ]

# use upliftRF to apply a Random Forest (alternatively use upliftKNN() to apply kNN). up.fit <- upliftRF(MOVED_AD_NUM ~ AGE + NH_WHITE + COMM_PT + H_F1 + REG_DAYS+

PR_PELIG + E_PELIG + POLITICALC + trt(MESSAGE_A), data = train.df, mtry = 3, ntree = 100, split_method = "KL", minsplit = 200, verbose = TRUE)

pred <- predict(up.fit, newdata = valid.df) # first colunm: p(y | treatment) # second colunm: p(y | control) head(data.frame(pred, "uplift" = pred[,1] - pred[,2]))

Output

pr.y1_ct1 pr.y1_ct0 uplift 1 0.356284 0.319376 0.036908 2 0.489685 0.437061 0.052624 3 0.380408 0.365436 0.014972 4 0.379597 0.341727 0.037870 5 0.371465 0.293659 0.077806 6 0.405424 0.349785 0.055639

Uplift modeling is used mainly in marketing and, more recently, in political campaigns. It has two main purposes:

• To determine whether to send someone a persuasion message, or just leave them alone.

• When a message is definitely going to be sent, to determine which mes- sage, among several possibilities, to send.

Technically this amounts to the same thing—“send no message” is simply another category of treatment, and an experiment can be constructed with mul- tiple treatments, for example, no message, message A, and message B. However, practitioners tend to think of the two purposes as distinct, and tend to focus on the first. Marketers want to avoid sending discount offers to customers who would make a purchase anyway, or renew a subscription anyway. Political cam- paigns, likewise, want to avoid calling voters who would vote for their candidate in any case. And both parties especially want to avoid sending messages or offers where the effect might be antagonistic—where the uplift is negative.

324 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

13.3 Summary

In practice, the methods discussed in this book are often used not in isolation, but as building blocks in an analytic process whose goal is always to inform and provide insight.

In this chapter, we looked at two ways that multiple models are deployed. In ensembles, multiple models are weighted and combined to produce improved predictions. In uplift modeling, the results of A-B testing are folded into the predictive modeling process as a predictor variable to guide choices not just about whether to send an offer or persuasion message, but also as to who to send it to.

PROBLEMS 325

PROBLEMS

13.1 Acceptance of Consumer Loan Universal Bank has begun a program to encourage its existing customers to borrow via a consumer loan program. The bank has promoted the loan to 5000 customers, of whom 480 accepted the offer. The data are available in file UniversalBank.csv. The bank now wants to develop a model to predict which customers have the greatest probability of accepting the loan, to reduce promotion costs and send the offer only to a subset of its customers.

We will develop several models, then combine them in an ensemble. The models we will use are (1) logistic regression, (2) k-nearest neighbors with k = 3, and (3) classification trees. Preprocess the data as follows:

• Zip code can be ignored.

• Partition the data: 60% training, 40% validation.

a. Fit models to the data for (1) logistic regression, (2) k-nearest neighbors with k = 3, and (3) classification trees. Use Personal Loan as the outcome variable. Report the validation confusion matrix for each of the three models.

b. Create a data frame with the actual outcome, predicted outcome, and each of the three models. Report the first 10 rows of this data frame.

c. Add two columns to this data frame for (1) a majority vote of predicted outcomes, and (2) the average of the predicted probabilities. Using the classifications generated by these two methods derive a confusion matrix for each method and report the overall accuracy.

d. Compare the error rates for the three individual methods and the two ensemble methods.

13.2 eBay Auctions—Boosting and Bagging Using the eBay auction data (file eBayAuc- tions.csv) with variable Competitive as the outcome variable, partition the data into training (60%) and validation (40%).

a. Run a classification tree, using the default controls of rpart(). Looking at the vali- dation set, what is the overall accuracy? What is the lift on the first decile?

b. Run a boosted tree with the same predictors (use function boosting() in the adabag package). For the validation set, what is the overall accuracy? What is the lift on the first decile?

c. Run a bagged tree with the same predictors (use function bagging() in the adabag package). For the validation set, what is the overall accuracy? What is the lift on the first decile?

d. Run a random forest (use function randomForest() in package randomForest with argument mtry = 4). Compare the bagged tree to the random forest in terms of validation accuracy and lift on first decile. How are the two methods conceptually different?

13.3 Predicting Delayed Flights (Boosting). The file FlightDelays.csv contains informa- tion on all commercial flights departing the Washington, DC area and arriving at New York during January 2004. For each flight there is information on the departure and arrival airports, the distance of the route, the scheduled time and date of the flight, and so on. The variable that we are trying to predict is whether or not a flight is delayed. A delay is defined as an arrival that is at least 15 minutes later than scheduled.

326 COMBINING METHODS: ENSEMBLES AND UPLIFT MODELING

Data Preprocessing. Transform variable day of week info a categorical variable. Bin the scheduled departure time into eight bins (in R use function cut()). Partition the data into training and validation sets.

Run a boosted classification tree for delay. Leave the default number of weak learners, and select resampling. Set maximum levels to display at 6, and minimum number of records in a terminal node to 1.

a. Compared with the single tree, how does the boosted tree behave in terms of overall accuracy?

b. Compared with the single tree, how does the boosted tree behave in terms of accuracy in identifying delayed flights?

c. Explain why this model might have the best performance over the other models you fit.

13.4 Hair Care Product—Uplift Modeling This problem uses the data set in Hair-Care- Product.csv, courtesy of SAS. In this hypothetical case, a promotion for a hair care product was sent to some members of a buyers club. Purchases were then recorded for both the members who got the promotion and those who did not.

a. What is the purchase propensity

i. among those who received the promotion?

ii. among those who did not receive the promotion?

b. Partition the data into training (60%) and validation (40%) and fit:

i. Uplift using a Random Forest.

ii. Uplift using k-NN.

c. Report the two models’ recommendations for the first three members.

Part V

Mining Relationships Among Records

CHAPTER 14

Association Rules and Collaborative Filtering

In this chapter, we describe the unsupervised learning methods of association rules (also called “affinity analysis” and “market basket analysis”) and collabora- tive filtering. Both methods are popular in marketing for cross-selling products associated with an item that a consumer is considering.

In association rules, the goal is to identify item clusters in transaction-type databases. Association rule discovery in marketing is termed “market basket anal- ysis” and is aimed at discovering which groups of products tend to be purchased together. These items can then be displayed together, offered in post-transaction coupons, or recommended in online shopping. We describe the two-stage pro- cess of rule generation and then assessment of rule strength to choose a subset. We look at the popular rule-generating Apriori algorithm, and then criteria for judging the strength of rules.

In collaborative filtering, the goal is to provide personalized recommen- dations that leverage user-level information. User-based collaborative filtering starts with a user, then finds users who have purchased a similar set of items or ranked items in a similar fashion, and makes a recommendation to the initial user based on what the similar users purchased or liked. Item-based collaborative fil- tering starts with an item being considered by a user, then locates other items that tend to be co-purchased with that first item. We explain the technique and the requirements for applying it in practice.

14.1 Association Rules

Put simply, association rules, or affinity analysis, constitute a study of “what goes with what.” This method is also called market basket analysis because it originated

329

330 ASSOCIATION RULES AND COLLABORATIVE FILTERING

with the study of customer transactions databases to determine dependencies between purchases of different items. Association rules are heavily used in retail for learning about items that are purchased together, but they are also useful in other fields. For example, a medical researcher might want to learn what symptoms appear together. In law, word combinations that appear too often might indicate plagiarism.

Discovering Association Rules in Transaction Databases

The availability of detailed information on customer transactions has led to the development of techniques that automatically look for associations between items that are stored in the database. An example is data collected using bar-code scanners in supermarkets. Such market basket databases consist of a large number of transaction records. Each record lists all items bought by a customer on a single-purchase transaction. Managers are interested to know if certain groups of items are consistently purchased together. They could use such information for making decisions on store layouts and item placement, for cross-selling, for promotions, for catalog design, and for identifying customer segments based on buying patterns. Association rules provide information of this type in the form of “if–then” statements. These rules are computed from the data; unlike the if–then rules of logic, association rules are probabilistic in nature.

Association rules are commonly encountered in online recommendation sys- tems (or recommender systems), where customers examining an item or items for possible purchase are shown other items that are often purchased in conjunction with the first item(s). The display from Amazon.com’s online shopping system illustrates the application of rules like this under “Frequently bought together.” In the example shown in Figure 14.1, a user browsing a Samsung Galaxy S5 cell phone is shown a case and a screen protector that are often purchased along with this phone.

We introduce a simple artificial example and use it throughout the chapter to demonstrate the concepts, computations, and steps of association rules. We end by applying association rules to a more realistic example of book purchases.

Example 1: Synthetic Data on Purchases of Phone Faceplates

A store that sells accessories for cellular phones runs a promotion on faceplates. Customers who purchase multiple faceplates from a choice of six different colors get a discount. The store managers, who would like to know what colors of faceplates customers are likely to purchase together, collected the transaction database as shown in Table 14.1.

Generating Candidate Rules

The idea behind association rules is to examine all possible rules between items in an if–then format, and select only those that are most likely to be indicators

ASSOCIATION RULES 331

FIGURE 14.1 RECOMMENDATIONS UNDER “FREQUENTLY BOUGHT TOGETHER” ARE BASED ON ASSOCIATION RULES

of true dependence. We use the term antecedent to describe the IF part, and consequent to describe the THEN part. In association analysis, the antecedent and consequent are sets of items (called itemsets) that are disjoint (do not have any items in common). Note that itemsets are not records of what people buy; they are simply possible combinations of items, including single items.

332 ASSOCIATION RULES AND COLLABORATIVE FILTERING

TABLE 14.1 TRANSACTIONS DATABASE FOR PURCHASES OF DIFFERENT-COLORED CELLULAR PHONE FACEPLATES

Transaction Faceplate Colors Purchased

1 red white green 2 white orange 3 white blue 4 red white orange 5 red blue 6 white blue 7 red blue 8 red white blue green 9 red white blue 10 yellow

Returning to the phone faceplate purchase example, one example of a pos- sible rule is “if red, then white,” meaning that if a red faceplate is purchased, a white one is, too. Here the antecedent is red and the consequent is white. The antecedent and consequent each contain a single item in this case. Another pos- sible rule is “if red and white, then green.” Here the antecedent includes the itemset {red, white} and the consequent is {green}.

The first step in association rules is to generate all the rules that would be candidates for indicating associations between items. Ideally, we might want to look at all possible combinations of items in a database with p distinct items (in the phone faceplate example, p = 6). This means finding all combinations of single items, pairs of items, triplets of items, and so on, in the transactions database. However, generating all these combinations requires a long computa- tion time that grows exponentially1 in p. A practical solution is to consider only combinations that occur with higher frequency in the database. These are called frequent itemsets.

Determining what qualifies as a frequent itemset is related to the concept of support. The support of a rule is simply the number of transactions that include both the antecedent and consequent itemsets. It is called a support because it measures the degree to which the data “support” the validity of the rule. The support is sometimes expressed as a percentage of the total number of records in the database. For example, the support for the itemset {red,white} in the phone faceplate example is 4 (or, 100× 4

10 = 40%).

What constitutes a frequent itemset is therefore defined as an itemset that has a support that exceeds a selected minimum support, determined by the user.

1The number of rules that one can generate for p items is 3p−2p+1+1. Computation time therefore grows by a factor for each additional item. For 6 items we have 602 rules, while for 7 items the number of rules grows to 1932.

ASSOCIATION RULES 333

The Apriori Algorithm

Several algorithms have been proposed for generating frequent itemsets, but the classic algorithm is the Apriori algorithm of Agrawal et al. (1993). The key idea of the algorithm is to begin by generating frequent itemsets with just one item (one-itemsets) and to recursively generate frequent itemsets with two items, then with three items, and so on, until we have generated frequent itemsets of all sizes.

It is easy to generate frequent one-itemsets. All we need to do is to count, for each item, how many transactions in the database include the item. These transaction counts are the supports for the one-itemsets. We drop one-itemsets that have support below the desired minimum support to create a list of the frequent one-itemsets.

To generate frequent two-itemsets, we use the frequent one-itemsets. The reasoning is that if a certain one-itemset did not exceed the minimum support, any larger size itemset that includes it will not exceed the minimum support. In general, generating k-itemsets uses the frequent (k − 1)-itemsets that were generated in the preceding step. Each step requires a single run through the database, and therefore the Apriori algorithm is very fast even for a large number of unique items in a database.

Selecting Strong Rules

From the abundance of rules generated, the goal is to find only the rules that indicate a strong dependence between the antecedent and consequent itemsets. To measure the strength of association implied by a rule, we use the measures of confidence and lift ratio, as described below.

Support and Confidence In addition to support, which we described earlier, there is another measure that expresses the degree of uncertainty about the if–then rule. This is known as the confidence2 of the rule. This measure compares the co-occurrence of the antecedent and consequent itemsets in the database to the occurrence of the antecedent itemsets. Confidence is defined as the ratio of the number of transactions that include all antecedent and consequent itemsets (namely, the support) to the number of transactions that include all the antecedent itemsets:

Confidence = no. transactions with both antecedent and consequent itemsets

no. transactions with antecedent itemset .

2The concept of confidence is different from and unrelated to the ideas of confidence intervals and confidence levels used in statistical inference.

334 ASSOCIATION RULES AND COLLABORATIVE FILTERING

For example, suppose that a supermarket database has 100,000 point-of-sale transactions. Of these transactions, 2000 include both orange juice and (over- the-counter) flu medication, and 800 of these include soup purchases. The association rule “IF orange juice and flu medication are purchased THEN soup is purchased on the same trip” has a support of 800 transactions (alternatively, 0.8% = 800/100,000) and a confidence of 40% (= 800/2000).

To see the relationship between support and confidence, let us think about what each is measuring (estimating). One way to think of support is that it is the (estimated) probability that a transaction selected randomly from the database will contain all items in the antecedent and the consequent:

Support = P̂ (antecedent AND consequent).

In comparison, the confidence is the (estimated) conditional probability that a trans- action selected randomly will include all the items in the consequent given that the transaction includes all the items in the antecedent:

Confidence= P̂ (antecedent AND consequent)

P̂ (antecedent) = P̂ (consequent | antecedent).

A high value of confidence suggests a strong association rule (in which we are highly confident). However, this can be deceptive because if the antecedent and/or the consequent has a high level of support, we can have a high value for confidence even when the antecedent and consequent are independent! For example, if nearly all customers buy bananas and nearly all customers buy ice cream, the confidence level of a rule such as “IF bananas THEN ice-cream” will be high regardless of whether there is an association between the items.

Lift Ratio A better way to judge the strength of an association rule is to compare the confidence of the rule with a benchmark value, where we assume that the occurrence of the consequent itemset in a transaction is independent of the occurrence of the antecedent for each rule. In other words, if the antecedent and consequent itemsets are independent, what confidence values would we expect to see? Under independence, the support would be

P (antecedent AND consequent) = P (antecedent) × P (consequent),

and the benchmark confidence would be

P (antecedent) × P (consequent) P (antecedent)

= P (consequent).

The estimate of this benchmark from the data, called the benchmark confidence value for a rule, is computed by

Benchmark confidence = no. transactions with consequent itemset

no. transactions in database .

ASSOCIATION RULES 335

We compare the confidence to the benchmark confidence by looking at their ratio: this is called the lift ratio of a rule. The lift ratio is the confidence of the rule divided by the confidence, assuming independence of consequent from antecedent:

lift ratio = confidence

benchmark confidence .

A lift ratio greater than 1.0 suggests that there is some usefulness to the rule. In other words, the level of association between the antecedent and consequent itemsets is higher than would be expected if they were independent. The larger the lift ratio, the greater the strength of the association.

To illustrate the computation of support, confidence, and lift ratio for the cellular phone faceplate example, we introduce an alternative presentation of the data that is better suited to this purpose.

Data Format

Transaction data are usually displayed in one of two formats: a transactions database (with each row representing a list of items purchased in a single trans- action), or a binary incidence matrix in which columns are items, rows again represent transactions, and each cell has either a 1 or a 0, indicating the presence or absence of an item in the transaction. For example, Table 14.1 displays the data for the cellular faceplate purchases in a transactions database. We translate these into binary incidence matrix format in Table 14.2.

Now suppose that we want association rules between items for this database that have a support count of at least 2 (equivalent to a percentage support of 2/10 = 20%): In other words, rules based on items that were purchased together in at least 20% of the transactions. By enumeration, we can see that only the itemsets listed in Table 14.3 have a count of at least 2.

The first itemset {red} has a support of 6, because six of the transactions included a red faceplate. Similarly, the last itemset {red, white, green} has a

TABLE 14.2 PHONE FACEPLATE DATA IN BINARY INCIDENCE MATRIX FORMAT

Transaction Red White Blue Orange Green Yellow

1 1 1 0 0 1 0 2 0 1 0 1 0 0 3 0 1 1 0 0 0 4 1 1 0 1 0 0 5 1 0 1 0 0 0 6 0 1 1 0 0 0 7 1 0 1 0 0 0 8 1 1 1 0 1 0 9 1 1 1 0 0 0

10 0 0 0 0 0 1

336 ASSOCIATION RULES AND COLLABORATIVE FILTERING

TABLE 14.3 ITEMSETS WITH SUPPORT COUNT OF AT LEAST TWO

Itemset Support (Count)

{red} 6 {white} 7 {blue} 6 {orange} 2 {green} 2 {red, white} 4 {red, blue} 4 {red, green} 2 {white, blue} 4 {white, orange} 2 {white, green} 2 {red, white, blue} 2 {red, white, green} 2

support of 2, because only two transactions included red, white, and green face- plates.

In R, the user will input data using the package arules in the transactions database format. The package only creates rules with one item as the consequent. It calls the consequent the right-hand side (RHS) of the rule, and the antecedent the left- hand-side (LHS) of the rule.

The Process of Rule Selection

The process of selecting strong rules is based on generating all association rules that meet stipulated support and confidence requirements. This is done in two stages. The first stage, described earlier, consists of finding all “frequent” item- sets, those itemsets that have a requisite support. In the second stage, we gener- ate, from the frequent itemsets, association rules that meet a confidence require- ment. The first step is aimed at removing item combinations that are rare in the database. The second stage then filters the remaining rules and selects only those with high confidence. For most association analysis data, the computa- tional challenge is the first stage, as described in the discussion of the Apriori algorithm.

The computation of confidence in the second stage is simple. Since any subset (e.g., {red} in the phone faceplate example) must occur at least as fre- quently as the set it belongs to (e.g., {red, white}), each subset will also be in the list. It is then straightforward to compute the confidence as the ratio of the support for the itemset to the support for each subset of the itemset. We retain the corresponding association rule only if it exceeds the desired cutoff value for confidence. For example, from the itemset {red, white, green} in the phone

ASSOCIATION RULES 337

faceplate purchases, we get the following single-consequent association rules, confidence values, and lift values:

Rule Confidence Lift

{red, white} ⇒ {green} support of {red, white, green}

support of {red, white} = 2/4 = 50%

confidence of rule benchmark confidence

= 50%

20% = 2.5

{green} ⇒ {red} support of {green, red}

support of {green} = 2/2 = 100%

confidence of rule benchmark confidence

= 100%

60% = 1.67

{white, green} ⇒ {red} support of {white, green, red}

support of {white, green} = 2/2 = 100%

confidence of rule benchmark confidence

= 100%

60% = 1.67

If the desired minimum confidence is 70%, we would report only the second and third rules.

We can generate association rules in R by coercing the binary incidence matrix in Table 14.2 into a transaction database like in Table 14.1. We specify the minimum support (20%) and minimum confidence level percentage (50%). Table 14.4 shows the output. The output includes information on each rule and its support, confidence, and lift (Note that here we consider all possible itemsets, not just {red, white, green} as above.).

Interpreting the Results

We can translate each of the rules from Table 14.4 into an understandable sen- tence that provides information about performance. For example, we can read rule #4 as follows:

If orange is purchased, then with confidence 100% white will also be purchased. This rule has a lift ratio of 1.43.

In interpreting results, it is useful to look at the various measures. The support for the rule indicates its impact in terms of overall size: How many transactions are affected? If only a small number of transactions are affected, the rule may be of little use (unless the consequent is very valuable and/or the rule is very efficient in finding it).

The lift ratio indicates how efficient the rule is in finding consequents, com- pared to random selection. A very efficient rule is preferred to an inefficient rule, but we must still consider support: A very efficient rule that has very low support may not be as desirable as a less efficient rule with much greater support.

The confidence tells us at what rate consequents will be found, and is useful in determining the business or operational usefulness of a rule: A rule with low confidence may find consequents at too low a rate to be worth the cost of (say) promoting the consequent in all the transactions that involve the antecedent.

338 ASSOCIATION RULES AND COLLABORATIVE FILTERING

TABLE 14.4 BINARY INCIDENCE MATRIX, TRANSACTIONS DATABASE, AND RULES FOR FACEPLATE EXAMPLE

code for running the Apriori algorithm

fp.df <- read.csv("Faceplate.csv")

# remove first column and convert to matrix fp.mat <- as.matrix(fp.df[, -1])

# convert the binary incidence matrix into a transactions database fp.trans <- as(fp.mat, "transactions") inspect(fp.trans)

## get rules # when running apriori(), include the minimum support, minimum confidence, and target # as arguments. rules <- apriori(fp.trans, parameter = list(supp = 0.2, conf = 0.5, target = "rules"))

# inspect the first six rules, sorted by their lift inspect(head(sort(rules, by = "lift"), n = 6))

Output

> fp.mat Red White Blue Orange Green Yellow

[1,] 1 1 0 0 1 0 [2,] 0 1 0 1 0 0 [3,] 0 1 1 0 0 0 [4,] 1 1 0 1 0 0 [5,] 1 0 1 0 0 0 [6,] 0 1 1 0 0 0 [7,] 1 0 1 0 0 0 [8,] 1 1 1 0 1 0 [9,] 1 1 1 0 0 0 [10,] 0 0 0 0 0 1

> inspect(fp.trans) items

1 {Red,White,Green} 2 {White,Orange} 3 {White,Blue} 4 {Red,White,Orange} 5 {Red,Blue} 6 {White,Blue} 7 {Red,Blue} 8 {Red,White,Blue,Green} 9 {Red,White,Blue} 10 {Yellow}

> inspect(head(sort(rules, by = "lift"), n = 6)) lhs rhs support confidence lift

15 {Red,White} => {Green} 0.2 0.5 2.500000 5 {Green} => {Red} 0.2 1.0 1.666667 14 {White,Green} => {Red} 0.2 1.0 1.666667 4 {Orange} => {White} 0.2 1.0 1.428571 6 {Green} => {White} 0.2 1.0 1.428571 13 {Red,Green} => {White} 0.2 1.0 1.428571

ASSOCIATION RULES 339

Rules and Chance

What about confidence in the nontechnical sense? How sure can we be that the rules we develop are meaningful? Considering the matter from a statistical perspective, we can ask: Are we finding associations that are really just chance occurrences?

Let us examine the output from an application of this algorithm to a small database of 50 transactions, where each of the nine items is assigned randomly to each transaction. The data are shown in Table 14.5, and the association rules generated are shown in Table 14.6. In looking at these tables, remember that “lhs” and “rhs” refer to itemsets, not records.

In this example, the lift ratios highlight Rule [105] as most interesting, as it suggests that purchase of item 4 is almost five times as likely when items 3 and 8

TABLE 14.5 FIFTY TRANSACTIONS OF RANDOMLY ASSIGNED ITEMS

Transaction Items Transaction Items Transaction Items

1 8 18 8 35 3 4 6 8 2 3 4 8 19 36 1 4 8 3 8 20 9 37 4 7 8 4 3 9 21 2 5 6 8 38 8 9 5 9 22 4 6 9 39 4 5 7 9 6 1 8 23 4 9 40 2 8 9 7 6 9 24 8 9 41 2 5 9 8 3 5 7 9 25 6 8 42 1 2 7 9 9 8 26 1 6 8 43 5 8 10 27 5 8 44 1 7 8 11 1 7 9 28 4 8 9 45 8 12 1 4 5 8 9 29 9 46 2 7 9 13 5 7 9 30 8 47 4 6 9 14 6 7 8 31 1 5 8 48 9 15 3 7 9 32 3 6 9 49 9 16 1 4 9 33 7 9 50 6 7 8 17 6 7 8 34 7 8 9

TABLE 14.6 ASSOCIATION RULES OUTPUT FOR RANDOM DATA

Min. Support: 2 = 4% Min. Conf. % : 70

> rules.tbl[rules.tbl$support >= 0.04 & rules.tbl$confidence >= 0.7,] lhs rhs support confidence lift

[18] {item.2} => {item.9} 0.08 0.8 1.481481 [89] {item.2,item.7} => {item.9} 0.04 1.0 1.851852 [104] {item.3,item.4} => {item.8} 0.04 1.0 1.851852 [105] {item.3,item.8} => {item.4} 0.04 1.0 5.000000 [113] {item.3,item.7} => {item.9} 0.04 1.0 1.851852 [119] {item.1,item.5} => {item.8} 0.04 1.0 1.851852 [149] {item.4,item.5} => {item.9} 0.04 1.0 1.851852 [155] {item.5,item.7} => {item.9} 0.06 1.0 1.851852 [176] {item.6,item.7} => {item.8} 0.06 1.0 1.851852

R code is provided in the next example

340 ASSOCIATION RULES AND COLLABORATIVE FILTERING

are purchased than if item 4 was not associated with the itemset {3,8}. Yet we know there is no fundamental association underlying these data—they were generated randomly.

Two principles can guide us in assessing rules for possible spuriousness due to chance effects:

1. The more records the rule is based on, the more solid the conclusion.

2. The more distinct rules we consider seriously (perhaps consolidating mul- tiple rules that deal with the same items), the more likely it is that at least some will be based on chance sampling results. For one person to toss a coin 10 times and get 10 heads would be quite surprising. If 1000 people toss a coin 10 times each, it would not be nearly so surprising to have one get 10 heads. Formal adjustment of “statistical significance” when multiple comparisons are made is a complex subject in its own right, and beyond the scope of this book. A reasonable approach is to consider rules from the top-down in terms of business or operational applicability, and not consider more than what can reasonably be incorporated in a human decision-making process. This will impose a rough constraint on the dangers that arise from an automated review of hundreds or thousands of rules in search of “something interesting.”

We now consider a more realistic example, using a larger database and real transactional data.

Example 2: Rules for Similar Book Purchases

The following example (drawn from the Charles Book Club case; see Chapter 21) examines associations among transactions involving various types of books. The database includes 2000 transactions, and there are 11 different types of books. The data, in binary incidence matrix form, are shown in Table 14.7.

TABLE 14.7 SUBSET OF BOOK PURCHASE TRANSACTIONS IN BINARY MATRIX FORMAT

ChildBks YouthBks CookBks DoItYBks cefBks ArtBks GeogBks ItalCook ItalAtlas ItalArt Florence

0 1 0 1 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 1 0 1 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 1 1 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

ASSOCIATION RULES 341

For instance, the first transaction included YouthBks (youth books) DoItYBks (do-it-yourself books), and GeogBks (geography books). Table 14.8 shows some of the rules generated for these data, given that we specified a minimal support of 200 transactions (out of 4000 transactions) and a minimal confidence of 50%. This resulted in 21 rules (see Table 14.8).

TABLE 14.8 RULES FOR BOOK PURCHASE TRANSACTIONS

code for running the Apriori algorithm

all.books.df <- read.csv("CharlesBookClub.csv")

# create a binary incidence matrix count.books.df <- all.books.df[, 8:18] incid.books.df <- ifelse(count.books.df > 0, 1, 0) incid.books.mat <- as.matrix(incid.books.df[, -1])

# convert the binary incidence matrix into a transactions database books.trans <- as(incid.books.mat, "transactions") inspect(books.trans)

# plot data itemFrequencyPlot(books.trans)

# run apriori function rules <- apriori(books.trans,

parameter = list(supp= 200/4000, conf = 0.5, target = "rules"))

# inspect rules inspect(sort(rules, by = "lift"))

Output

> inspect(sort(rules, by = "lift")) lhs rhs support confidence lift

16 {DoItYBks,GeogBks} => {YouthBks} 0.05450 0.5396040 2.264864 18 {CookBks,GeogBks} => {YouthBks} 0.08025 0.5136000 2.155719 13 {CookBks,RefBks} => {DoItYBks} 0.07450 0.5330948 2.092619 14 {YouthBks,GeogBks} => {DoItYBks} 0.05450 0.5215311 2.047227 20 {YouthBks,CookBks} => {DoItYBks} 0.08375 0.5201863 2.041948 10 {YouthBks,RefBks} => {CookBks} 0.06825 0.8400000 2.021661 15 {YouthBks,DoItYBks} => {GeogBks} 0.05450 0.5278450 1.978801 19 {YouthBks,DoItYBks} => {CookBks} 0.08375 0.8111380 1.952197 12 {DoItYBks,RefBks} => {CookBks} 0.07450 0.8054054 1.938400 11 {RefBks,GeogBks} => {CookBks} 0.06450 0.7889908 1.898895 17 {YouthBks,GeogBks} => {CookBks} 0.08025 0.7679426 1.848237 21 {DoItYBks,GeogBks} => {CookBks} 0.07750 0.7673267 1.846755 7 {YouthBks,ArtBks} => {CookBks} 0.05150 0.7410072 1.783411 9 {DoItYBks,ArtBks} => {CookBks} 0.05300 0.7114094 1.712177 3 {RefBks} => {CookBks} 0.13975 0.6825397 1.642695 8 {ArtBks,GeogBks} => {CookBks} 0.05525 0.6800000 1.636582 4 {YouthBks} => {CookBks} 0.16100 0.6757608 1.626380 6 {DoItYBks} => {CookBks} 0.16875 0.6624141 1.594258 1 {ItalCook} => {CookBks} 0.06875 0.6395349 1.539193 5 {GeogBks} => {CookBks} 0.15625 0.5857545 1.409758 2 {ArtBks} => {CookBks} 0.11300 0.5067265 1.219558

342 ASSOCIATION RULES AND COLLABORATIVE FILTERING

In reviewing these rules, we see that the information can be compressed. First, rule #1, which appears from the confidence level to be a very promis- ing rule, is probably meaningless. It says: “If Italian cooking books have been purchased, then cookbooks are purchased.” It seems likely that Italian cooking books are simply a subset of cookbooks. Rules 14,15, and 16 involve the same trio of books, with different antecedents and consequents. The same is true of rules 17 and 18 as well as 19 and 20. (Pairs and groups like this are easy to track down by looking for rows that share the same support.) This does not mean that the rules are not useful. On the contrary, it can reduce the number of itemsets to be considered for possible action from a business perspective.

14.2 Collaborative Filtering3

Recommendation systems are a critically important part of websites that offer a large variety of products or services. Examples include Amazon.com, which offers millions of different products; Netflix has thousands of movies for rental; Google searches over huge numbers of webpages; Internet radio websites such as Spotify and Pandora include a large variety of music albums by various artists; travel websites offer many destinations and hotels; social network websites have many groups. The recommender engine provides personalized recommenda- tions to a user based on the user’s information as well as on similar users’ infor- mation. Information means behaviors indicative of preference, such as purchase, ratings, and clicking.

The value that recommendation systems provide to users helps online com- panies convert browsers into buyers, increase cross-selling, and build loyalty.

Collaborative filtering is a popular technique used by such recommendation systems. The term collaborative filtering is based on the notions of identifying relevant items for a specific user from the very large set of items (“filtering”) by considering preferences of many users (“collaboration”).

The Fortune.com article “Amazon’s Recommendation Secret” (June 30, 2012) describes the company’s use of collaborative filtering not only for provid- ing personalized product recommendations, but also for customizing the entire website interface for each user:

At root, the retail giant’s recommendation system is based on a number of simple elements: what a user has bought in the past, which items they have in their virtual shopping cart, items they’ve rated and liked, and what other customers have viewed and purchased. Amazon calls this homegrown math “item-to-item collaborative filtering,” and it’s used this algorithm to heavily customize the browsing experience for returning customers.

3This section copyright © 2017 Datastats, LLC, Galit Shmueli, and Peter Bruce.

COLLABORATIVE FILTERING 343

Data Type and Format

Collaborative filtering requires availability of all item–user information. Specif- ically, for each item–user combination, we should have some measure of the user’s preference for that item. Preference can be a numerical rating or a binary behavior such as a purchase, a ‘like’, or a click.

For n users (u1, u2, . . . , un) and p items (i1, i2, . . . ip), we can think of the data as an n×p matrix of n rows (users) by p columns (items). Each cell includes the rating or the binary event corresponding to the user’s preference of the item (see schematic in Table 14.9). Typically not every user purchases or rates every item, and therefore a purchase matrix will have many zeros (it is sparse), and a rating matrix will have many missing values. Such missing values sometimes convey “uninterested” (as opposed to non-missing values that convey interest).

When both n and p are large, it is not practical to store the preferences data (ru,i) in an n× p table. Instead, the data can be stored in many rows of triplets of the form (Uu, Ii, ru,i), where each triplet contains the user ID, the item ID, and the preference information.

Item ID User ID I1 I2 · · · Ip U1 r1,1 r1,2 · · · r1,p U2 r2,1 r2,2 · · · r2,p ...

Un rn,1 rn,2 · · · rn,p

TABLE 14.9 SCHEMATIC OF MATRIX FORMAT WITH RATINGS DATA

Example 3: Netflix Prize Contest

We have been considering both association rules and collaborative filtering as unsupervised techniques, but it is possible to judge how well they do by looking at holdout data to see what users purchase and how they rate items. The famous Netflix contest, mentioned in Chapter 13, did just this and provides a useful example to illustrate collaborative filtering, though the extension into training and validation is beyond the scope of this book.

In 2006, Netflix, the largest movie rental service in North America, announced a one million USD contest (www.netflixprize.com) for the purpose of improving its recommendation system called Cinematch. Participants were provided with a number of datasets, one for each movie. Each dataset included all the customer ratings for that movie (and the timestamp). We can think of one large combined dataset of the form [customer ID, movie ID, rating, date] where each record includes the rating given by a certain customer to a certain movie

344 ASSOCIATION RULES AND COLLABORATIVE FILTERING

TABLE 14.10 SAMPLE OF RECORDS FROM THE NETFLIX PRIZE CONTEST, FOR A SUBSET OF 10 CUSTOMERS AND 9 MOVIES

Movie ID Customer ID 1 5 8 17 18 28 30 44 48

30878 4 1 3 3 4 5 124105 4 822109 5 823519 3 1 4 4 5 885013 4 5 893988 3 4 4

1248029 3 2 4 3 1503895 4 1842128 4 3 2238063 3

on a certain date. Ratings were on a 1–5 star scale. Contestants were asked to develop a recommendation algorithm that would improve over the existing Net- flix system. Table 14.10 shows a small sample from the contest data, organized in matrix format. Rows indicate customers and columns are different movies.

It is interesting to note that the winning team was able to improve their system by considering not just the ratings for a movie, but whether a movie was rated by a particular customer or not. In other words, the information on which movies a customer decided to rate turned out to be critically informative of cus- tomers’ preferences, more than simply considering the 1–5 rating information4:

Collaborative filtering methods address the sparse set of rating values. However, much accuracy is obtained by also looking at other features of the data. First is the information on which movies each user chose to rate, regardless of specific rating value (“the binary view”). This played a decisive role in our 2007 solution, and reflects the fact that the movies to be rated are selected deliberately by the user, and are not a random sample.

This is an example where converting the rating information into a binary matrix of rated/unrated proved to be useful.

User-Based Collaborative Filtering: “People Like You”

One approach to generating personalized recommendations for a user using col- laborative filtering is based on finding users with similar preferences, and recom- mending items that they liked but the user hasn’t purchased. The algorithm has two steps:

4Bell, R. M., Koren, Y., and Volinsky, C., “The BellKor 2008 Solution to the Netflix Prize”, www. netflixprize.com/assets/ProgressPrize2008_BellKor.pdf.

COLLABORATIVE FILTERING 345

1. Find users who are most similar to the user of interest (neighbors). This is done by comparing the preference of our user to the preferences of other users.

2. Considering only the items that the user has not yet purchased, recom- mend the ones that are most preferred by the user’s neighbors.

This is the approach behind Amazon’s “Customers Who Bought This Item Also Bought…” (see Figure 14.1). It is also used in a Google search for generating the “Similar pages” link shown near each search result.

Step 1 requires choosing a distance (or proximity) metric to measure the distance between our user and the other users. Once the distances are computed, we can use a threshold on the distance or on the number of required neighbors to determine the nearest neighbors to be used in Step 2. This approach is called “user-based top-N recommendation.”

A nearest-neighbors approach measures the distance of our user to each of the other users in the database, similar to the k-nearest-neighbors algorithm (see Chapter 7). The Euclidean distance measure we discussed in that chap- ter does not perform as well for collaborative filtering as some other mea- sures. A popular proximity measure between two users is the Pearson correlation between their ratings. We denote the ratings of items I1, . . . , Ip by user U1 as r1,1, r1,2, . . . , r1,p and their average by r1. Similarly, the ratings by user U2 are r2,1, r2,2, . . . , r2,p, with average r2. The correlation proximity between the two users is defined by

Corr(U1, U2) =

∑ (r1,i − r1)(r2,i − r2)√∑

(r1,i − r1)2 √∑

(r2,i − r2)2 , (14.1)

where the summations are only over the items co-rated by both users. To illustrate this, let us compute the correlation between customer 30878

and customer 823519 in the small Netflix sample in Table 14.10. We’ll assume that the data shown in the table is the entire information. First, we compute the average rating by each of these users:

r30878 = (4 + 1 + 3 + 3 + 4 + 5)/6 = 3.333

r823519 = (3 + 1 + 4 + 4 + 5)/5 = 3.4

Note that the average is computed over a different number of movies for each of these customers, because they each rated a different set of movies. The average for a customer is computed over all the movies that a customer rated. The calculations for the correlation involve the departures from the average, but

346 ASSOCIATION RULES AND COLLABORATIVE FILTERING

only for the items that they co-rated. In this case the co-rated movie IDs are 1, 28, and 30:

Corr(U30878, U823519) =

(4− 3.333)(3− 3.4) + (3− 3.333)(4− 3.4) + (4− 3.333)(5− 3.4)√ (4− 3.333)2 + (3− 3.333)2 + (4− 3.333)2

√ (3− 3.4)2 + (4− 3.4)2 + (5− 3.4)2

= 0.6/1.75 = 0.34

The same approach can be used when the data are in the form of a binary matrix (e.g., purchased or didn’t purchase.)

Another popular measure is a variant of the Pearson correlation called cosine similarity. It differs from the correlation formula by not subtracting the means. Subtracting the mean in the correlation formula adjusts for users’ different overall approaches to rating—for example, a customer who always rates highly vs. one who tends to give low ratings.5

For example, the cosine similarity between the two Netflix customers is:

Cos Sim(U30878, U823519) = 4× 3 + 3× 4 + 4× 5√

42 + 32 + 42 √ 32 + 42 + 52

= 44/45.277 = 0.972

Note that when the data are in the form of a binary matrix, say, for purchase or no-purchase, the cosine similarity must be calculated over all items that either user has purchased; it cannot be limited to just the items that were co-purchased.

Collaborative filtering suffers from what is called a cold start: it cannot be used as is to create recommendations for new users or new items. For a user who rated a single item, the correlation coefficient between this and other users (in user-generated collaborative filtering) will have a denominator of zero and the cosine proximity will be 1 regardless of the rating. In a similar vein, users with just one item, and items with just one user, do not qualify as candidates for nearby neighbors.

For a user of interest, we compute his/her similarity to each of the users in our database using a correlation, cosine similarity, or another measure. Then, in Step 2, we look only at the k nearest users, and among all the other items that they rated/purchased, we choose the best one and recommend it to our user. What is the best one? For binary purchase data, it is the item most purchased.

5Correlation and cosine similarity are popular in collaborative filtering because they are computation- ally fast for high-dimensional sparse data, and they account both for the rating values and the number of rated items.

COLLABORATIVE FILTERING 347

For rating data, it could be the highest rated, most rated, or a weighting of the two.

The nearest-neighbors approach can be computationally expensive when we have a large database of users. One solution is to use clustering methods (see Chapter 15) to group users into homogeneous clusters in terms of their preferences, and then to measure the distance of our user to each of the clusters. This approach places the computational load on the clustering step that can take place earlier and offline; it is then cheaper (and faster) to compare our user to each of the clusters in real time. The price of clustering is less accurate recommendations, because not all the members of the closest cluster are the most similar to our user.

Item-Based Collaborative Filtering

When the number of users is much larger than the number of items, it is com- putationally cheaper (and faster) to find similar items rather than similar users. Specifically, when a user expresses interest in a particular item, the item-based collaborative filtering algorithm has two steps:

1. Find the items that were co-rated, or co-purchased, (by any user) with the item of interest.

2. Recommend the most popular or correlated item(s) among the similar items.

Similarity is now computed between items, instead of users. For example, in our small Netflix sample (Table 14.10), the correlation between movie 1 (with average r1 = 3.7) and movie 5 (with average r5 = 3) is:

Corr(I1, I5) = (4− 3.7)(1− 3) + (4− 3.7)(5− 3)√

(4− 3.7)2 + (4− 3.7)2 √ (1− 3)2 + (5− 3)2

= 0

The zero correlation is due to the two opposite ratings of movie 5 by the users who also rated 1. One user rated it 5 stars and the other gave it a 1 star.

In like fashion, we can compute similarity between all the movies. This can be done offline. In real time, for a user who rates a certain movie highly, we can look up the movie correlation table and recommend the movie with the highest positive correlation to the user’s newly rated movie.

According to an industry report6 by researchers who developed the Amazon item-to-item recommendation system,

“[The item-based] algorithm produces recommendations in real time, scales to massive data sets, and generates high-quality recommendations.”

6“Linden, G., Smith, B., and York J., Amazon.com Recommendations: Item-to-Item Collaborative Filtering”, IEEE Internet Computing, vol. 7, no. 1, p. 76–80, 2003.

348 ASSOCIATION RULES AND COLLABORATIVE FILTERING

The disadvantage of item-based recommendations is that there is less diversity between items (compared to users’ taste), and therefore, the recommendations are often obvious.

Table 14.11 shows R code for collaborative filtering on a simulated data similar to the Netflix data using the recommenderlab package.

TABLE 14.11 COLLABORATIVE FILTERING IN R

code for collaborative filtering

# simulate matrix with 1000 users and 100 movies m <- matrix(nrow = 1000, ncol = 100) # simulated ratings (1% of the data) m[sample.int(100*1000, 1000)] <- ceiling(runif(1000, 0, 5)) ## convert into a realRatingMatrix r <- as(m, "realRatingMatrix")

library(recommenderlab) # user-based collaborative filtering UB.Rec <- Recommender(r, "UBCF") pred <- predict(UB.Rec, r, type="ratings") as(pred, "matrix")

# item-based collaborative filtering IB.Rec <- Recommender(r, "IBCF") pred <- predict(IB.Rec, r, type="ratings") as(pred, "matrix")

Advantages and Weaknesses of Collaborative Filtering

Collaborative filtering relies on the availability of subjective information regard- ing users’ preferences. It provides useful recommendations, even for “long tail” items, if our database contains sufficient similar users (not necessarily many, but at least a few per user), so that each user can find other users with similar tastes. Similarly, the data should include sufficient per-item ratings or purchases. One limitation of collaborative filtering is therefore that it cannot generate recom- mendations for new users, nor for new items. There are various approaches for tackling this challenge.

User-based collaborative filtering looks for similarity in terms of highly-rated or preferred items. However, it is blind to data on low-rated or unwanted items. We can therefore not expect to use it as is for detecting unwanted items.

User-based collaborative filtering helps leverage similarities between peo- ple’s tastes for providing personalized recommendations. However, when the number of users becomes very large, collaborative filtering becomes computa- tionally difficult. Solutions include item-based algorithms, clustering of users,

COLLABORATIVE FILTERING 349

and dimension reduction. The most popular dimension reduction method used in such cases is singular value decomposition (SVD), a computationally superior form of principal components analysis (see Chapter 4).

Although the term “prediction” is often used to describe the output of col- laborative filtering, this method is unsupervised by nature. It can be used to generate predicted ratings or purchase indication for a user, but usually we do not have the true outcome value in practice. One important way to improve rec- ommendations generated by collaborative filtering is by getting user feedback. Once a recommendation is generated, the user can indicate whether the recom- mendation was adequate or not. For this reason, many recommender systems entice users to provide feedback on their recommendations.

Collaborative Filtering vs. Association Rules

While collaborative filtering and association rules are both unsupervised methods used for generating recommendations, they differ in several ways:

Frequent itemsets vs. personalized recommendations Association rules look for frequent item combinations and will provide recommendations only for those items. In contrast, collaborative filtering provides personalized recommendations for every item, thereby catering to users with unusual taste. In this sense, collaborative filtering is useful for capturing the “long tail” of user preferences, while association rules look for the “head.” This difference has implications for the data needed: association rules require data on a very large number of “baskets” (transactions) in order to find a sufficient number of baskets that contain certain combinations of items. In contrast, collaborative filtering does not require many “baskets,” but does require data on as many items as possible for many users. Also, association rules operate at the basket level (our database can include multiple transactions for each user), while collaborative filtering operates at the user level.

Because association rules produce generic, impersonal rules (association- based recommendations such as Amazon’s “Frequently Bought Together” display the same recommendations to all users searching for a specific item), they can be used for setting common strategies such as product placement in a store or sequencing of diagnostic tests in hospitals. In contrast, col- laborative filtering generates user-specific recommendations (e.g., Amazon’s “Customers Who Bought This Item Also Bought…”) and is therefore a tool designed for personalization.

Transactional data vs. user data Association rules provide recommen- dations of items based on their co-purchase with other items in many trans- actions/baskets. In contrast, collaborative filtering provides recommendations of items based on their co-purchase or co-rating by even a small number of

350 ASSOCIATION RULES AND COLLABORATIVE FILTERING

other users. Considering distinct baskets is useful when the same items are purchased over and over again (e.g., in grocery shopping). Considering dis- tinct users is useful when each item is typically purchased/rated once (e.g., purchases of books, music, and movies).

Binary data and ratings data Association rules treat items as binary data (1 = purchase, 0 = nonpurchase), whereas collaborative filtering can operate on either binary data or on numerical ratings.

Two or more items In association rules, the antecedent and consequent can each include one or more items (e.g., IF milk THEN cookies and corn- flakes). Hence, a recommendation might be a bundle of the item of interest with multiple items (“buy milk, cookies, and cornflakes and receive 10% dis- count”). In contrast, in collaborative filtering, similarity is measured between pairs of items or pairs of users. A recommendation will therefore be either for a single item (the most popular item purchased by people like you, which you haven’t purchased), or for multiple single items which do not necessarily relate to each other (the top two most popular items purchased by people like you, which you haven’t purchased).

These distinctions are sharper for purchases and recommendations of non- popular items, especially when comparing association rules to user-based collab- orative filtering. When considering what to recommend to a user who purchased a popular item, then association rules and item-based collaborative filtering might yield the same recommendation for a single item. But a user-based recommen- dation will likely differ. Consider a customer who purchases milk every week as well as gluten-free products (which are rarely purchased by other customers). Suppose that using association rules on the transaction database we identify the rule “IF milk THEN cookies.” Then, the next time our customer purchases milk, s/he will receive a recommendation (e.g., a coupon) to purchase cookies, whether or not s/he purchased cookies, and irrespective of his/her gluten-free item purchases. In item-based collaborative filtering, we would look at all items co-purchased with milk across all users and recommend the most popular item among them (which was not purchased by our customer). This might also lead to a recommendation of cookies, because this item was not purchased by our customer.7 Now consider user-based collaborative filtering. User-based collab- orative filtering searches for similar customers—those who purchased the same set of items—and then recommend the item most commonly purchased by these neighbors, which was not purchased by our customer. The user-based recommenda- tion is therefore unlikely to recommend cookies and more likely to recommend popular gluten-free items that the customer has not purchased.

7If the rule is “IF milk, THEN cookies and cornflakes” then the association rules would recommend cookies and cornflakes to a milk purchaser, while item-based collaborative filtering would recommend the most popular single item purchased with milk.

SUMMARY 351

14.3 Summary

Association rules (also called market basket analysis) and collaborative filtering are unsupervised methods for deducing associations between purchased items from databases of transactions. Association rules search for generic rules about items that are purchased together. The main advantage of this method is that it generates clear, simple rules of the form “IF X is purchased, THEN Y is also likely to be purchased.” The method is very transparent and easy to understand.

The process of creating association rules is two-staged. First, a set of candi- date rules based on frequent itemsets is generated (the Apriori algorithm being the most popular rule-generating algorithm). Then from these candidate rules, the rules that indicate the strongest association between items are selected. We use the measures of support and confidence to evaluate the uncertainty in a rule. The user also specifies minimal support and confidence values to be used in the rule generation and selection process. A third measure, the lift ratio, com- pares the efficiency of the rule to detect a real association compared to a random combination.

One shortcoming of association rules is the profusion of rules that are gen- erated. There is therefore a need for ways to reduce these to a small set of useful and strong rules. An important nonautomated method to condense the infor- mation involves examining the rules for uninformative and trivial rules as well as for rules that share the same support. Another issue that needs to be kept in mind is that rare combinations tend to be ignored, because they do not meet the minimum support requirement. For this reason, it is better to have items that are approximately equally frequent in the data. This can be achieved by using higher-level hierarchies as the items. An example is to use types of books rather than titles of individual books in deriving association rules from a database of bookstore transactions.

Collaborative filtering is a popular technique used in online recommenda- tion systems. It is based on the relationship between items formed by users who acted similarly on an item, such as purchasing or rating an item highly. User- based collaborative filtering operates on data on item–user combinations, calcu- lates the similarities between users and provides personalized recommendations to users. An important component for the success of collaborative filtering is that users provide feedback about the recommendations provided and have sufficient information on each item. One disadvantage of collaborative filtering methods is that they cannot generate recommendations for new users or new items. Also, with a huge number of users, user-based collaborative filtering becomes compu- tationally challenging, and alternatives such as item-based methods or dimension reduction are popularly used.

352 ASSOCIATION RULES AND COLLABORATIVE FILTERING

PROBLEMS

14.1 Satellite Radio Customers. An analyst at a subscription-based satellite radio com- pany has been given a sample of data from their customer database, with the goal of finding groups of customers who are associated with one another. The data consist of company data, together with purchased demographic data that are mapped to the company data (see Table 14.12). The analyst decides to apply association rules to learn more about the associations between customers. Comment on this approach.

TABLE 14.12 SAMPLE OF DATA ON SATELLITE RADIO CUSTOMERS

Row ID zipconvert_2 zipconvert_3 zipconvert_4 zipconvert_5 homeowner NUMCHLD INCOME gender WEALTH dummy dummy

17 0 1 0 0 1 1 5 1 9 25 1 0 0 0 1 1 1 0 7 29 0 0 0 1 0 2 5 1 8 38 0 0 0 1 1 1 3 0 4 40 0 1 0 0 1 1 4 0 8 53 0 1 0 0 1 1 4 1 8 58 0 0 0 1 1 1 4 1 8 61 1 0 0 0 1 1 1 0 7 71 0 0 1 0 1 1 4 0 5 87 1 0 0 0 1 1 4 1 8

100 0 0 0 1 1 1 4 1 8 104 1 0 0 0 1 1 1 1 5 121 0 0 1 0 1 1 4 1 5 142 1 0 0 0 0 1 5 0 8

14.2 Identifying Course Combinations. The Institute for Statistics Education at Statis- tics.com offers online courses in statistics and analytics, and is seeking information that will help in packaging and sequencing courses. Consider the data in the file Course- Topics.csv, the first few rows of which are shown in Table 14.13. These data are for purchases of online statistics courses at Statistics.com. Each row represents the courses attended by a single customer. The firm wishes to assess alternative sequencings and bundling of courses. Use association rules to analyze these data, and interpret several of the resulting rules.

TABLE 14.13 DATA ON PURCHASES OF ONLINE STATISTICS COURSES

Intro DataMining Survey CatData Regression Forecast DOE SW

1 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 1 1 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0

PROBLEMS 353

14.3 Recommending Courses We again consider the data in CourseTopics.csv describing course purchases at Statistics.com (see Problem 14.2 and data sample in Table 14.13). We want to provide a course recommendation to a student who purchased the Regres- sion and Forecast courses. Apply user-based collaborative filtering to the data. You will get a Null matrix. Explain why this happens.

14.4 Cosmetics Purchases. The data shown in Table 14.14 and the output in Table 14.15 are based on a subset of a dataset on cosmetic purchases (Cosmetics.csv) at a large chain

TABLE 14.14 EXCERPT FROM DATA ON COSMETICS PURCHASES IN BINARY MATRIX FORM

Trans. # Bag Blush Nail Polish Brushes Concealer Eyebrow Pencils Bronzer

1 0 1 1 1 1 0 1 2 0 0 1 0 1 0 1 3 0 1 0 0 1 1 1 4 0 0 1 1 1 0 1 5 0 1 0 0 1 0 1 6 0 0 0 0 1 0 0 7 0 1 1 1 1 0 1 8 0 0 1 1 0 0 1 9 0 0 0 0 1 0 0

10 1 1 1 1 0 0 0 11 0 0 1 0 0 0 1 12 0 0 1 1 1 0 1

TABLE 14.15 ASSOCIATION RULES FOR COSMETICS PURCHASES DATA

lhs rhs support confidence lift 1 {Blush,

Concealer, Mascara, Eye.shadow, Lipstick} => {Eyebrow.Pencils} 0.013 0.3023255814 7.198228128

2 {Trans., Blush, Concealer, Mascara, Eye.shadow, Lipstick} => {Eyebrow.Pencils} 0.013 0.3023255814 7.198228128

3 {Blush, Concealer, Mascara, Lipstick} => {Eyebrow.Pencils} 0.013 0.2888888889 6.878306878

4 {Trans., Blush, Concealer, Mascara, Lipstick} => {Eyebrow.Pencils} 0.013 0.2888888889 6.878306878

5 {Blush, Concealer, Eye.shadow, Lipstick} => {Eyebrow.Pencils} 0.013 0.2826086957 6.728778468

6 {Trans., Blush, Concealer, Eye.shadow, Lipstick} => {Eyebrow.Pencils} 0.013 0.2826086957 6.728778468

354 ASSOCIATION RULES AND COLLABORATIVE FILTERING

drugstore. The store wants to analyze associations among purchases of these items for purposes of point-of-sale display, guidance to sales personnel in promoting cross- sales, and guidance for piloting an eventual time-of-purchase electronic recommender system to boost cross-sales. Consider first only the data shown in Table 14.14, given in binary matrix form.

a. Select several values in the matrix and explain their meaning.

b. Consider the results of the association rules analysis shown in Table 14.15.

i. For the first row, explain the “confidence” output and how it is calculated.

ii. For the first row, explain the “support” output and how it is calculated.

iii. For the first row, explain the “lift” and how it is calculated.

iv. For the first row, explain the rule that is represented there in words.

c. Now, use the complete dataset on the cosmetics purchases (in the file Cosmetics.csv). Using R, apply association rules to these data (use the default parameters).

i. Interpret the first three rules in the output in words.

ii. Reviewing the first couple of dozen rules, comment on their redundancy and how you would assess their utility.

14.5 Course ratings. The Institute for Statistics Education at Statistics.com asks students to rate a variety of aspects of a course as soon as the student completes it. The Institute is contemplating instituting a recommendation system that would provide students with recommendations for additional courses as soon as they submit their rating for a completed course. Consider the excerpt from student ratings of online statistics courses shown in Table 14.16, and the problem of what to recommend to student E.N.

a. First consider a user-based collaborative filter. This requires computing correlations between all student pairs. For which students is it possible to compute correlations with E.N.? Compute them.

b. Based on the single nearest student to E.N., which single course should we recom- mend to E.N.? Explain why.

c. Use R (function similarity()) to compute the cosine similarity between users.

TABLE 14.16 RATINGS OF ONLINE STATISTICS COURSES: 4 = BEST, 1 = WORST, BLANK = NOT TAKEN

SQL Spatial PA 1 DM in R Python Forecast R Prog Hadoop Regression

L N 4 3 2 4 2 M H 3 4 4 J H 2 2 E N 4 4 4 3 D U 4 4 F L 4 G L 4 A H 3 S A 4 R W 2 4 B A 4 M G 4 4 A F 4 K G 3 D S 4 2 4

PROBLEMS 355

d. Based on the cosine similarities of the nearest students to E.N., which course should be recommended to E.N.?

e. What is the conceptual difference between using the correlation as opposed to cosine similarities? [Hint: how are the missing values in the matrix handled in each case?]

f. With large datasets, it is computationally difficult to compute user-based recom- mendations in real time, and an item-based approach is used instead. Returning to the rating data (not the binary matrix), let’s now take that approach.

i. If the goal is still to find a recommendation for E.N., for which course pairs is it possible and useful to calculate correlations?

ii. Just looking at the data, and without yet calculating course pair correlations, which course would you recommend to E.N., relying on item-based filter- ing? Calculate two course pair correlations involving your guess and report the results.

g. Apply item-based collaborative filtering to this dataset (using R) and based on the results, recommend a course to E.N.

CHAPTER 15

Cluster Analysis

This chapter is about the popular unsupervised learning task of clustering, where the goal is to segment the data into a set of homogeneous clusters of records for the purpose of generating insight. Separating a dataset into clusters of homoge- neous records is also useful for improving performance of supervised methods, by modeling each cluster separately rather than the entire, heterogeneous dataset. Clustering is used in a vast variety of business applications, from customized marketing to industry analysis. We describe two popular clustering approaches: hierarchical clustering and k-means clustering. In hierarchical clustering, records are sequentially grouped to create clusters, based on distances between records and distances between clusters. We describe how the algorithm works in terms of the clustering process and mention several common distance metrics used. Hierar- chical clustering also produces a useful graphical display of the clustering process and results, called a dendrogram. We present dendrograms and illustrate their usefulness. k-means clustering is widely used in large dataset applications. In k-means clustering, records are allocated to one of a prespecified set of clusters, according to their distance from each cluster. We describe the k-means cluster- ing algorithm and its computational advantages. Finally, we present techniques that assist in generating insight from clustering results.

15.1 Introduction

Cluster analysis is used to form groups or clusters of similar records based on several measurements made on these records. The key idea is to characterize the clusters in ways that would be useful for the aims of the analysis. This idea has been applied in many areas, including astronomy, archaeology, medicine, chem- istry, education, psychology, linguistics, and sociology. Biologists, for example,

357

358 CLUSTER ANALYSIS

have made extensive use of classes and subclasses to organize species. A spectac- ular success of the clustering idea in chemistry was Mendeleev’s periodic table of the elements.

One popular use of cluster analysis in marketing is for market segmentation: customers are segmented based on demographic and transaction history infor- mation, and a marketing strategy is tailored for each segment. In countries such as India, where customer diversity is extremely location-sensitive, chain stores often perform market segmentation at the store level, rather than chain-wide (called “micro segmentation”). Another use is for market structure analysis: iden- tifying groups of similar products according to competitive measures of similar- ity. In marketing and political forecasting, clustering of neighborhoods using US postal zip codes has been used successfully to group neighborhoods by lifestyles. Claritas, a company that pioneered this approach, grouped neighborhoods into 40 clusters using various measures of consumer expenditure and demographics. Examining the clusters enabled Claritas to come up with evocative names, such as “Bohemian Mix,” “Furs and Station Wagons,” and “Money and Brains,” for the groups that captured the dominant lifestyles. Knowledge of lifestyles can be used to estimate the potential demand for products (such as sports utility vehicles) and services (such as pleasure cruises). Similarly, sales organizations will derive customer segments and give them names—“personas”—to focus sales efforts.

In finance, cluster analysis can be used for creating balanced portfolios: Given data on a variety of investment opportunities (e.g., stocks), one may find clus- ters based on financial performance variables such as return (daily, weekly, or monthly), volatility, beta, and other characteristics, such as industry and market capitalization. Selecting securities from different clusters can help create a bal- anced portfolio. Another application of cluster analysis in finance is for industry analysis: For a given industry, we are interested in finding groups of similar firms based on measures such as growth rate, profitability, market size, product range, and presence in various international markets. These groups can then be ana- lyzed in order to understand industry structure and to determine, for instance, who is a competitor.

An interesting and unusual application of cluster analysis, described in Berry and Linoff (1997), is the design of a new set of sizes for army uniforms for women in the US Army. The study came up with a new clothing size system with only 20 sizes, where different sizes fit different body types. The 20 sizes are combinations of five measurements: chest, neck, and shoulder circumference, sleeve outseam, and neck-to-buttock length (for further details, see McCullugh et al., 1998). This example is important because it shows how a completely new insightful view can be gained by examining clusters of records.

Cluster analysis can be applied to huge amounts of data. For instance, Inter- net search engines use clustering techniques to cluster queries that users submit. These can then be used for improving search algorithms. The objective of this

INTRODUCTION 359

chapter is to describe the key ideas underlying the most commonly used tech- niques for cluster analysis and to lay out their strengths and weaknesses.

Typically, the basic data used to form clusters are a table of measurements on several variables, where each column represents a variable and a row represents a record. Our goal is to form groups of records so that similar records are in the same group. The number of clusters may be prespecified or determined from the data.

Example: Public Utilities

Table 15.1 gives corporate data on 22 public utilities in the United States (the variable definitions are given in the table footnote). We are interested in forming groups of similar utilities. The records to be clustered are the utilities, and the clustering will be based on the eight measurements on each utility. An example

TABLE 15.1 DATA ON 22 PUBLIC UTILITIES

Company Fixed RoR Cost Load Demand Sales Nuclear Fuel Cost

Arizona Public Service 1.06 9.2 151 54.4 1.6 9077 0.0 0.628 Boston Edison Co. 0.89 10.3 202 57.9 2.2 5088 25.3 1.555 Central Louisiana Co. 1.43 15.4 113 53.0 3.4 9212 0.0 1.058 Commonwealth Edison Co. 1.02 11.2 168 56.0 0.3 6423 34.3 0.700 Consolidated Edison Co. (NY) 1.49 8.8 192 51.2 1.0 3300 15.6 2.044 Florida Power & Light Co. 1.32 13.5 111 60.0 −2.2 11127 22.5 1.241 Hawaiian Electric Co. 1.22 12.2 175 67.6 2.2 7642 0.0 1.652 daho Power Co. 1.10 9.2 245 57.0 3.3 13082 0.0 0.309 Kentucky Utilities Co. 1.34 13.0 168 60.4 7.2 8406 0.0 0.862 Madison Gas & Electric Co. 1.12 12.4 197 53.0 2.7 6455 39.2 0.623 Nevada Power Co. 0.75 7.5 173 51.5 6.5 17441 0.0 0.768 New England Electric Co. 1.13 10.9 178 62.0 3.7 6154 0.0 1.897 Northern States Power Co. 1.15 12.7 199 53.7 6.4 7179 50.2 0.527 Oklahoma Gas & Electric Co. 1.09 12.0 96 49.8 1.4 9673 0.0 0.588 Pacific Gas & Electric Co. 0.96 7.6 164 62.2 −0.1 6468 0.9 1.400 Puget Sound Power & Light Co. 1.16 9.9 252 56.0 9.2 15991 0.0 0.620 San Diego Gas & Electric Co. 0.76 6.4 136 61.9 9.0 5714 8.3 1.920 The Southern Co. 1.05 12.6 150 56.7 2.7 10140 0.0 1.108 Texas Utilities Co. 1.16 11.7 104 54.0 −2.1 13507 0.0 0.636 Wisconsin Electric Power Co. 1.20 11.8 148 59.9 3.5 7287 41.1 0.702 United Illuminating Co. 1.04 8.6 204 61.0 3.5 6650 0.0 2.116 Virginia Electric & Power Co. 1.07 9.3 174 54.3 5.9 10093 26.6 1.306

Fixed = fixed-charge covering ratio (income/debt); RoR = rate of return on capital Cost = cost per kilowatt capacity in place; Load = annual load factor Demand = peak kilowatthour demand growth from 1974 to 1975 Sales = sales (kilowatthour use per year) Nuclear = percent nuclear Fuel Cost = total fuel costs (cents per kilowatthour)

360 CLUSTER ANALYSIS

FIGURE 15.1 SCATTER PLOT OF FUEL COST VS. SALES FOR THE 22 UTILITIES

where clustering would be useful is a study to predict the cost impact of dereg- ulation. To do the requisite analysis, economists would need to build a detailed cost model of the various utilities. It would save a considerable amount of time and effort if we could cluster similar types of utilities and build detailed cost models for just one “typical” utility in each cluster and then scale up from these models to estimate results for all utilities.

For simplicity, let us consider only two of the measurements: Sales and Fuel Cost. Figure 15.1 shows a scatter plot of these two variables, with labels marking each company. At first glance, there appear to be two or three clusters of utilities: one with utilities that have high fuel costs, a second with utilities that have lower fuel costs and relatively low sales, and a third with utilities with low fuel costs but high sales.

We can therefore think of cluster analysis as a more formal algorithm that measures the distance between records, and according to these distances (here, two-dimensional distances), forms clusters.

Two general types of clustering algorithms for a dataset of n records are hierarchical and non-hierarchical clustering:

Hierarchical methods can be either agglomerative or divisive. Agglomera- tive methods begin with n clusters and sequentially merge similar clusters until a single cluster is obtained. Divisive methods work in the opposite direction, starting with one cluster that includes all records. Hierarchical methods are especially useful when the goal is to arrange the clusters into a natural hierarchy.

MEASURING DISTANCE BETWEEN TWO RECORDS 361

Non-hierarchical methods, such as k-means. Using a prespecified num- ber of clusters, the method assigns records to each cluster. These methods are generally less computationally intensive and are therefore preferred with very large datasets.

We concentrate here on the two most popular methods: hierarchical agglomerative clustering and k-means clustering. In both cases, we need to define two types of distances: distance between two records and distance between two clusters. In both cases, there is a variety of metrics that can be used.

15.2 Measuring Distance Between Two Records

We denote by dij a distance metric, or dissimilarity measure, between records i and j. For record i we have the vector of p measurements (xi1, xi2, . . . , xip), while for record j we have the vector of measurements (xj1, xj2, . . . , xjp). For example, we can write the measurement vector for Arizona Public Service as [1.06, 9.2, 151, 54.4, 1.6, 9077, 0, 0.628].

Distances can be defined in multiple ways, but in general, the following properties are required:

Non-negative: dij ≥ 0 Self-proximity: dii = 0 (the distance from a record to itself is zero)

Symmetry: dij = dji

Triangle inequality: dij ≤ dik + dkj (the distance between any pair can- not exceed the sum of distances between the other two pairs)

Euclidean Distance

The most popular distance measure is the Euclidean distance, dij , which between two records, i and j, is defined by

dij = √ (xi1 − xj1)2 + (xi2 − xj2)2 + · · ·+ (xip − xjp)2.

For instance, the Euclidean distance between Arizona Public Service and Boston Edison Co. can be computed from the raw data by

d12 = √

(1.06− 0.89)2 + (9.2− 10.3)2 + (151− 202)2 + · · ·+ (0.628− 1.555)2 = 3989.408.

To compute Euclidean distance in R, see Table 15.2.

362 CLUSTER ANALYSIS

TABLE 15.2 DISTANCE MATRIX BETWEEN PAIRS OF THE FIRST FIVE UTILITIES, USING EUCLIDEAN DISTANCE

code for computing distance between records

utilities.df <- read.csv("Utilities.csv")

# set row names to the utilities column row.names(utilities.df) <- utilities.df[,1]

# remove the utility column utilities.df <- utilities.df[,-1]

# compute Euclidean distance # (to compute other distance measures, change the value in method = ) d <- dist(utilities.df, method = "euclidean")

Output

> d Arizona Boston Central Commonwealth NY

Boston 3989.40808 Central 140.40286 4125.04413 Commonwealth 2654.27763 1335.46650 2789.75967 NY 5777.16767 1788.06803 5912.55291 3123.15322 Florida 2050.52944 6039.68908 1915.15515 4704.36310 7827.42921

Normalizing Numerical Measurements

The measure computed above is highly influenced by the scale of each vari- able, so that variables with larger scales (e.g., Sales) have a much greater influ- ence over the total distance. It is therefore customary to normalize continuous measurements before computing the Euclidean distance. This converts all mea- surements to the same scale. Normalizing a measurement means subtracting the average and dividing by the standard deviation (normalized values are also called z-scores). For instance, the average sales amount across the 22 utilities is 8914.045 and the standard deviation is 3549.984. The normalized sales for Arizona Public Service is therefore (9077− 8914.045)/3549.984 = 0.046.

Returning to the simplified utilities data with only two measurements (Sales and Fuel Cost), we first normalize the measurements (see Table 15.3), and then compute the Euclidean distance between each pair. Table 15.4 gives these pair- wise distances for the first five utilities. A similar table can be constructed for all 22 utilities.

Other Distance Measures for Numerical Data

It is important to note that the choice of the distance measure plays a major role in cluster analysis. The main guideline is domain dependent: What exactly is being

MEASURING DISTANCE BETWEEN TWO RECORDS 363

TABLE 15.3 ORIGINAL AND NORMALIZED MEASUREMENTS FOR SALES AND FUEL COST

Company Sales Fuel Cost NormSales NormFuel

Arizona Public Service 9077 0.628 0.0459 −0.8537 Boston Edison Co. 5088 1.555 −1.0778 0.8133 Central Louisiana Co. 9212 1.058 0.0839 −0.0804 Commonwealth Edison Co. 6423 0.7 −0.7017 −0.7242 Consolidated Edison Co. (NY) 3300 2.044 −1.5814 1.6926 Florida Power & Light Co. 11127 1.241 0.6234 0.2486 Hawaiian Electric Co. 7642 1.652 −0.3583 0.9877 Idaho Power Co. 13082 0.309 1.1741 −1.4273 Kentucky Utilities Co. 8406 0.862 −0.1431 −0.4329 Madison Gas & Electric Co. 6455 0.623 −0.6927 −0.8627 Nevada Power Co. 17441 0.768 2.4020 −0.6019 New England Electric Co. 6154 1.897 −0.7775 1.4283 Northern States Power Co. 7179 0.527 −0.4887 −1.0353 Oklahoma Gas & Electric Co. 9673 0.588 0.2138 −0.9256 Pacific Gas & Electric Co. 6468 1.4 −0.6890 0.5346 Puget Sound Power & Light Co. 15991 0.62 1.9935 −0.8681 San Diego Gas & Electric Co. 5714 1.92 −0.9014 1.4697 The Southern Co. 10140 1.108 0.3453 0.0095 Texas Utilities Co. 13507 0.636 1.2938 −0.8393 Wisconsin Electric Power Co. 7287 0.702 −0.4583 −0.7206 United Illuminating Co. 6650 2.116 −0.6378 1.8221 Virginia Electric & Power Co. 10093 1.306 0.3321 0.3655

Mean 8914.05 1.10 0.00 0.00 Standard deviation 3549.98 0.56 1.00 1.00

TABLE 15.4 DISTANCE MATRIX BETWEEN PAIRS OF THE FIRST FIVE UTILITIES, USING EUCLIDEAN DISTANCE AND NORMALIZED MEASUREMENTS

code for normalizing data and computing distance

# normalize input variables utilities.df.norm <- sapply(utilities.df, scale)

# add row names: utilities row.names(utilities.df.norm) <- row.names(utilities.df)

# compute normalized distance based on Sales (column 6) and Fuel Cost (column 8) d.norm <- dist(utilities.df.norm[,c(6,8)], method = "euclidean")

Output

> d.norm Arizona Boston Central Commonwealth NY

Boston 2.0103293 Central 0.7741795 1.4657027 Commonwealth 0.7587375 1.5828208 1.0157104 NY 3.0219066 1.0133700 2.4325285 2.5719693 Florida 1.2444219 1.7923968 0.6318918 1.6438566 2.6355728

364 CLUSTER ANALYSIS

measured? How are the different measurements related? What scale should each measurement be treated as (numerical, ordinal, or nominal)? Are there outliers? Finally, depending on the goal of the analysis, should the clusters be distinguished mostly by a small set of measurements, or should they be separated by multiple measurements that weight moderately?

Although Euclidean distance is the most widely used distance, it has three main features that need to be kept in mind. First, as mentioned earlier, it is highly scale dependent. Changing the units of one variable (e.g., from cents to dollars) can have a huge influence on the results. Normalizing is therefore a common solution. But unequal weighting should be considered if we want the clusters to depend more on certain measurements and less on others. The second feature of Euclidean distance is that it completely ignores the relation- ship between the measurements. Thus, if the measurements are in fact strongly correlated, a different distance (such as the statistical distance, described later) is likely to be a better choice. Third, Euclidean distance is sensitive to outliers. If the data are believed to contain outliers and careful removal is not a choice, the use of more robust distances (such as the Manhattan distance, described later) is preferred.

Additional popular distance metrics often used (for reasons such as the ones above) are:

Correlation-based similarity. Sometimes it is more natural or conve- nient to work with a similarity measure between records rather than distance, which measures dissimilarity. A popular similarity measure is the square of the Pearson correlation coefficient, r2ij , where the correlation coefficient is defined by

rij =

p∑ m=1

(xim − xm)(xjm − xm)√ p∑

m=1 (xim − xm)2

p∑ m=1

(xjm − xm)2 . (15.1)

Such measures can always be converted to distance measures. In the example above, we could define a distance measure dij = 1− r2ij .

Statistical distance (also called Mahalanobis distance). This metric has an advantage over the other metrics mentioned in that it takes into account the correlation between measurements. With this metric, measurements that are highly correlated with other measurements do not contribute as much as those that are uncorrelated or mildly correlated. The statistical distance between records i and j is defined as

di,j = √ (xi − xj)′S−1(xi − xj),

MEASURING DISTANCE BETWEEN TWO RECORDS 365

where xi and xj are p-dimensional vectors of the measurements values for records i and j, respectively; and S is the covariance matrix for these vec- tors. (′, a transpose operation, simply turns a column vector into a row vector). S−1 is the inverse matrix of S, which is the p-dimension extension to division. For further information on statistical distance, see Chapter 12.

Manhattan distance (“city block”). This distance looks at the absolute differences rather than squared differences, and is defined by

dij =

p∑ m=1

| xim − xjm | .

Maximum coordinate distance. This distance looks only at the mea- surement on which records i and j deviate most. It is defined by

dij = max m=1,2,...,p

| xim − xjm | .

Distance Measures for Categorical Data

In the case of measurements with binary values, it is more intuitively appealing to use similarity measures than distance measures. Suppose that we have binary values for all the xij ’s, and for records i and j we have the following 2× 2 table:

Record j

0 1

Record i 0 a b a+ b

1 c d c+ d

a+ c b+ d n

where a denotes the number of variables for which records i and j do not have that attribute (they each have value 0 on that attribute), d is the number of variables for which the two records have the attribute present, and so on. The most useful similarity measures in this situation are:

Matching coefficient: (a+ d)/n.

Jaquard’s coefficient: d/(b+c+d). This coefficient ignores zero matches. This is desirable when we do not want to consider two people to be similar simply because a large number of characteristics are absent in both. For example, if owns a Corvette is one of the variables, a matching “yes” would be evidence of similarity, but a matching “no” tells us little about whether the two people are similar.

366 CLUSTER ANALYSIS

Distance Measures for Mixed Data

When the measurements are mixed (some continuous and some binary), a simi- larity coefficient suggested by Gower is very useful. Gower’s similarity measure is a weighted average of the distances computed for each variable, after scaling each variable to a [0,1] scale. It is defined as

sij =

p∑ m=1

wijmsijm

p∑ m=1

wijm

,

where sijm is the similarity between records i and j on measurement m, and wijm is a binary weight given to the corresponding distance.

The similarity measures sijm and weights wijm are computed as follows:

1. For continuous measurements, sijm = 1− |xim−xjm|max(xm)−min(xm) and wijm = 1 unless the value for measurement m is unknown for one or both of the records, in which case wijm = 0.

2. For binary measurements, sijm = 1 if xim = xjm = 1 and 0 otherwise. wijm = 1 unless xim = xjm = 0.

3. For nonbinary categorical measurements, sijm = 1 if both records are in the same category, and otherwise sijm = 0. As in continuous measure- ments, wijm = 1 unless the category for measurement m is unknown for one or both of the records, in which case wijm = 0.

15.3 Measuring Distance Between Two Clusters

We define a cluster as a set of one or more records. How do we measure dis- tance between clusters? The idea is to extend measures of distance between records into distances between clusters. Consider cluster A, which includes the m records A1, A2, . . . , Am and cluster B, which includes n records B1, B2, . . . , Bn. The most widely used measures of distance between clusters are:

Minimum Distance

The distance between the pair of records Ai and Bj that are closest:

min(distance(Ai, Bj)), i = 1, 2, . . . ,m; j = 1, 2, . . . , n.

Maximum Distance

The distance between the pair of records Ai and Bj that are farthest:

max(distance(Ai, Bj)), i = 1, 2, . . . ,m; j = 1, 2, . . . , n.

MEASURING DISTANCE BETWEEN TWO CLUSTERS 367

Average Distance

The average distance of all possible distances between records in one cluster and records in the other cluster:

Average(distance(Ai, Bj)), i = 1, 2, . . . ,m; j = 1, 2, . . . , n.

Centroid Distance

The distance between the two cluster centroids. A cluster centroid is the vector of measurement averages across all the records in that cluster. For cluster A, this is the vector xA = [(1/m

∑m i=1 x1i, . . . , 1/m

∑m i=1 xpi)]. The centroid distance

between clusters A and B is

distance(xA, xB).

Minimum distance, maximum distance, and centroid distance are illustrated visually for two dimensions with a map of Portugal and France in Figure 15.2.

For instance, consider the first two utilities (Arizona, Boston) as cluster A, and the next three utilities (Central, Commonwealth, Consolidated) as cluster B. Using the normalized scores in Table 15.3 and the distance matrix in Table

FIGURE 15.2 TWO-DIMENSIONAL REPRESENTATION OF SEVERAL DIFFERENT DISTANCE MEASURES BETWEEN PORTUGAL AND FRANCE

368 CLUSTER ANALYSIS

15.4, we can compute each of the distances described above. Using Euclidean distance for each distance calculation, we get:

• The closest pair is Arizona and Commonwealth, and therefore the mini- mum distance between clusters A and B is 0.76.

• The farthest pair is Arizona and Consolidated, and therefore the maxi- mum distance between clusters A and B is 3.02.

• The average distance is (0.77+0.76+3.02+1.47+1.58+1.01)/6 = 1.44.

• The centroid of cluster A is[ 0.0459− 1.0778

2 , −0.8537 + 0.8133

2

] = [−0.516,−0.020],

and the centroid of cluster B is[ 0.0839− 0.7017− 1.5814

3 , −0.0804− 0.7242 + 1.6926

3

] = [−0.733, 0.296].

The distance between the two centroids is then√ (−0.516 + 0.733)2 + (−0.020− 0.296)2 = 0.38.

In deciding among clustering methods, domain knowledge is key. If you have good reason to believe that the clusters might be chain- or sausage-like, minimum distance would be a good choice. This method does not require that cluster members all be close to one another, only that the new members being added be close to one of the existing members. An example of an application where this might be the case would be characteristics of crops planted in long rows, or disease outbreaks along navigable waterways that are the main areas of settlement in a region. Another example is laying and finding mines (land or marine). Minimum distance is also fairly robust to small deviations in the distances. However, adding or removing data can influence it greatly.

Maximum and average distance are better choices if you know that the clus- ters are more likely to be spherical (e.g., customers clustered on the basis of numerous attributes). If you do not know the probable nature of the cluster, these are good default choices, since most clusters tend to be spherical in nature.

We now move to a more detailed description of the two major types of clustering algorithms: hierarchical (agglomerative) and non-hierarchical.

15.4 Hierarchical (Agglomerative) Clustering

The idea behind hierarchical agglomerative clustering is to start with each cluster comprising exactly one record and then progressively agglomerating (combining)

HIERARCHICAL (AGGLOMERATIVE) CLUSTERING 369

the two nearest clusters until there is just one cluster left at the end, which consists of all the records.

Returning to the small example of five utilities and two measures (Sales and Fuel Cost) and using the distance matrix (Table 15.4), the first step in the hier- archical clustering would join Arizona and Commonwealth, which are the clos- est (using normalized measurements and Euclidean distance). Next, we would recalculate a 4× 4 distance matrix that would have the distances between these four clusters: {Arizona, Commonwealth}, {Boston}, {Central}, and {Consol- idated}. At this point, we use a measure of distance between clusters such as the ones described in Section 15.3. Each of these distances (minimum, maximum, average, and centroid distance) can be implemented in the hierarchical scheme as described below.

H I E R A R C H I C A L A G G L O M E R A T I V E C L U S T E R I N G A L G O R I T H M :

1. Start with n clusters (each record = cluster).

2. The two closest records are merged into one cluster.

3. At every step, the two clusters with the smallest distance are merged. This means that either single records are added to existing clusters or two existing clusters are combined.

Single Linkage

In single linkage clustering, the distance measure that we use is the minimum dis- tance (the distance between the nearest pair of records in the two clusters, one record in each cluster). In our utilities example, we would compute the dis- tances between each of {Boston}, {Central}, and {Consolidated} with {Ari- zona, Commonwealth} to create the 4×4 distance matrix shown in Table 15.5.

The next step would consolidate {Central} with {Arizona, Common- wealth} because these two clusters are closest. The distance matrix will again be recomputed (this time it will be 3× 3), and so on.

This method has a tendency to cluster together at an early stage records that are distant from each other because of a chain of intermediate records in the

TABLE 15.5 DISTANCE MATRIX AFTER ARIZONA AND COMMONWEALTH CONSOLIDATION CLUSTER TOGETHER, USING SINGLE LINKAGE

Arizona–Commonwealth Boston Central Consolidated

Arizona–Commonwealth 0 Boston min(2.01,1.58) 0 Central min(0.77,1.02) 1.47 0 Consolidated min(3.02,2.57) 1.01 2.43 0

370 CLUSTER ANALYSIS

same cluster. Such clusters have elongated sausage-like shapes when visualized as objects in space.

Complete Linkage

In complete linkage clustering, the distance between two clusters is the maximum distance (between the farthest pair of records). If we used complete linkage with the five-utilities example, the recomputed distance matrix would be equivalent to Table 15.5, except that the “min” function would be replaced with a “max.”

This method tends to produce clusters at the early stages with records that are within a narrow range of distances from each other. If we visualize them as objects in space, the records in such clusters would have roughly spherical shapes.

Average Linkage

Average linkage clustering is based on the average distance between clusters (between all possible pairs of records). If we used average linkage with the five- utilities example, the recomputed distance matrix would be equivalent to Table 15.5, except that the “min” function would be replaced with “average.” This method is also called Unweighted Pair-Group Method using Averages (UPGMA).

Note that unlike average linkage, the results of the single and complete link- age methods depend only on the ordering of the inter-record distances. Linear transformations of the distances (and other transformations that do not change the ordering) do not affect the results.

Centroid Linkage

Centroid linkage clustering is based on centroid distance, where clusters are rep- resented by their mean values for each variable, which forms a vector of means. The distance between two clusters is the distance between these two vectors. In average linkage, each pairwise distance is calculated, and the average of all such distances is calculated. In contrast, in centroid distance clustering, just one distance is calculated: the distance between group means. This method is also called Unweighted Pair-Group Method using Centroids (UPGMC).

Ward’s Method

Ward’s method is also agglomerative, in that it joins records and clusters together progressively to produce larger and larger clusters, but operates slightly differently from the general approach described above. Ward’s method considers the “loss of information” that occurs when records are clustered together. When each cluster has one record, there is no loss of information and all individual values remain available. When records are joined together and represented in clusters, information about an individual record is replaced by the information for the

HIERARCHICAL (AGGLOMERATIVE) CLUSTERING 371

cluster to which it belongs. To measure loss of information, Ward’s method employs a measure “error sum of squares” (ESS) that measures the difference between individual records and a group mean.

This is easiest to see in univariate data. For example, consider the values (2, 6, 5, 6, 2, 2, 2, 2, 0, 0, 0) with a mean of 2.5. Their ESS is equal to

(2− 2.5)2 + (6− 2.5)2 + (5− 2.5)2 + . . .+ (0− 2.5)2 = 50.5.

The loss of information associated with grouping the values into a single group is therefore 50.5. Now group the records into four groups: (0, 0, 0), (2, 2, 2, 2), (5), (6, 6). The loss of information is the sum of the ESS’s for each group, which is 0 (each record in each group is equal to the mean for that group, so the ESS for each group is 0). Thus, clustering the 10 records into 4 clusters results in no loss of information, and this would be the first step in Ward’s method. In moving to a smaller number of clusters, Ward’s method would choose the configuration that results in the smallest incremental loss of information.

Ward’s method tends to result in convex clusters that are of roughly equal size, which can be an important consideration in some applications (e.g., in establishing meaningful customer segments).

Dendrograms: Displaying Clustering Process and Results

A dendrogram is a treelike diagram that summarizes the process of clustering. On the x-axis are the records. Similar records are joined by lines whose ver- tical length reflects the distance between the records. Figure 15.3 shows the dendrograms that results from clustering the 22 utilities using the 8 normalized measurements, Euclidean distance, once with single linkage (top) and once with average linkage (bottom).

By choosing a cutoff distance on the y-axis, a set of clusters is created. Visu- ally, this means drawing a horizontal line on a dendrogram. Records with con- nections below the horizontal line (that is, their distance is smaller than the cutoff distance) belong to the same cluster. For example, setting the cutoff distance to 2.7 on the single linkage dendrogram in Figure 15.3 (top) results in six clusters. The six clusters are (from left to right on the dendrogram):

{NY}, {Nevada}, {San Diego}, {Idaho, Puget}, {Central}, {Others}.

If we want six clusters using average linkage, we can choose a cutoff distance of 3.5. The resulting six clusters are slightly different.

The six (or other number of) clusters can be computed in R by applying the function cutree() to the dendrogram object. Table 15.6 shows the results for the single linkage and the average linkage clustering with six clusters. Each record is assigned a cluster number. While some records remain in the same cluster

372 CLUSTER ANALYSIS

code for running hierarchical clustering and generating a dendrogram

# in hclust() set argument method = # to "ward.D", "single", "complete", "average", "median", or "centroid" hc1 <- hclust(d.norm, method = "single") plot(hc1, hang = -1, ann = FALSE) hc2 <- hclust(d.norm, method = "average") plot(hc2, hang = -1, ann = FALSE)

N Y

N ev

ad a

S an

D ie

go

Id ah

o

P ug

et

C en

tr al

F lo

rid a

V irg

in ia

P ac

ifi c

H aw

ai ia

n

N ew

E ng

la nd

U ni

te d

B os

to n

W is

co ns

in

C om

m on

w ea

lth

M ad

is on

N or

th er

n

K en

tu ck

y

O kl

ah om

a

Te xa

s

A riz

on a

S ou

th er

n

1. 0

1. 5

2. 0

2. 5

3. 0

3. 5

N ev

ad a

Id ah

o

P ug

et N Y

S an

D ie

go

P ac

ifi c

H aw

ai ia

n

N ew

E ng

la nd

U ni

te d

M ad

is on

N or

th er

n

C om

m on

w ea

lth

W is

co ns

in

B os

to n

V irg

in ia

C en

tr al

K en

tu ck

y

F lo

rid a

O kl

ah om

a

Te xa

s

A riz

on a

S ou

th er

n

1. 0

1. 5

2. 0

2. 5

3. 0

3. 5

4. 0

4. 5

FIGURE 15.3 DENDROGRAM: SINGLE LINKAGE (TOP) AND AVERAGE LINKAGE (BOTTOM) FOR ALL 22 UTILITIES, USING ALL EIGHT MEASUREMENTS

HIERARCHICAL (AGGLOMERATIVE) CLUSTERING 373

TABLE 15.6 COMPUTING CLUSTER MEMBERSHIP BY “CUTTING” THE DENDROGRAM

Single Linkage

> memb <- cutree(hc1, k = 6) > memb

Arizona Boston Central Commonwealth NY 1 1 2 1 3

Florida Hawaiian Idaho Kentucky Madison 1 1 4 1 1

Nevada New England Northern Oklahoma Pacific 5 1 1 1 1

Puget San Diego Southern Texas Wisconsin 4 6 1 1 1

United Virginia 1 1

Average Linkage

> memb <- cutree(hc2, k = 6) > memb

Arizona Boston Central Commonwealth NY 1 2 1 2 3

Florida Hawaiian Idaho Kentucky Madison 1 4 5 1 2

Nevada New England Northern Oklahoma Pacific 5 4 2 1 4

Puget San Diego Southern Texas Wisconsin 5 6 1 1 2

United Virginia 4 2

in both methods (e.g., Arizona, Florida, Kentucky, Oklamona, Texas), others change.

Note that if we wanted five clusters, they would be identical to the six above, with the exception that two of the clusters would be merged into a single cluster. For example, in the single linkage case, the two right-most clusters in the den- drogram would be merged into one cluster. In general, all hierarchical methods have clusters that are nested within each other as we decrease the number of clusters. This is a valuable property for interpreting clusters and is essential in certain applications, such as taxonomy of varieties of living organisms.

Validating Clusters

One important goal of cluster analysis is to come up withmeaningful clusters. Since there are many variations that can be chosen, it is important to make sure that the resulting clusters are valid, in the sense that they really generate some insight. To see whether the cluster analysis is useful, consider each of the following aspects:

1. Cluster interpretability. Is the interpretation of the resulting clusters reason- able? To interpret the clusters, explore the characteristics of each cluster by

374 CLUSTER ANALYSIS

a. Obtaining summary statistics (e.g., average, min, max) from each clus- ter on each measurement that was used in the cluster analysis

b. Examining the clusters for separation along some common feature (variable) that was not used in the cluster analysis

c. Labeling the clusters: based on the interpretation, trying to assign a name or label to each cluster

2. Cluster stability. Do cluster assignments change significantly if some of the inputs are altered slightly? Another way to check stability is to partition the data and see how well clusters formed based on one part apply to the other part. To do this:

a. Cluster partition A.

b. Use the cluster centroids from A to assign each record in partition B (each record is assigned to the cluster with the closest centroid).

c. Assess how consistent the cluster assignments are compared to the assignments based on all the data.

3. Cluster separation. Examine the ratio of between-cluster variation to within-cluster variation to see whether the separation is reasonable. There exist statistical tests for this task (an F-ratio), but their usefulness is somewhat controversial.

4. Number of clusters. The number of resulting clusters must be useful, given the purpose of the analysis. For example, suppose the goal of the clus- tering is to identify categories of customers and assign labels to them for market segmentation purposes. If the marketing department can only manage to sustain three different marketing presentations, it would prob- ably not make sense to identify more than three clusters.

Returning to the utilities example, we notice that both methods (single and average linkage) identify {NY} and {San Diego} as singleton clusters. Also, both dendrograms imply that a reasonable number of clusters in this dataset is four. One insight that can be derived from the average linkage clustering is that clusters tend to group geographically. The four non-singleton clusters form (approxi- mately) a southern group, a northern group, an east/west seaboard group, and a west group.

We can further characterize each of the clusters by examining the summary statistics of their measurements, or visually looking at a heatmap of their indi- vidual measurements. Figure 15.4 shows a heatmap of the four clusters and two singletons, highlighting the different profile that each cluster has in terms of the eight measurements. We see, for instance, that cluster 2 is characterized by utili- ties with a high percent of nuclear power; cluster 1 is characterized by high fixed charge and RoR; cluster 4 has high fuel costs.

HIERARCHICAL (AGGLOMERATIVE) CLUSTERING 375

code for creating heatmap

# set labels as cluster membership and utility name row.names(utilities.df.norm) <- paste(memb, ": ", row.names(utilities.df), sep = "")

# plot heatmap # rev() reverses the color mapping to large = dark heatmap(as.matrix(utilities.df.norm), Colv = NA, hclustfun = hclust,

col=rev(paste("gray",1:99,sep="")))

F ix

ed _c

ha rg

e

R oR

C os

t

Lo ad

_f ac

to r

D em

an d_

gr ow

th

S al

es

N uc

le ar

F ue

l_ C

os t

1: Oklahoma 5: Texas 1: Arizona 3: Southern 3: Florida 3: Central 1: Kentucky 4: NY 2: Boston 3: Virginia 1: Commonwealth 1: Wisconsin 1: Madison 1: Northern 4: San Diego 2: Pacific 4: United 4: New England 2: Hawaiian 6: Nevada 5: Idaho 6: Puget

FIGURE 15.4 HEATMAP FOR THE 22 UTILITIES (IN ROWS). ROWS ARE SORTED BY THE SIX CLUSTERS FROM AVERAGE LINKAGE CLUSTERING. DARKER CELLS DENOTE HIGHER VALUES WITHIN A COLUMN

Limitations of Hierarchical Clustering

Hierarchical clustering is very appealing in that it does not require specification of the number of clusters, and in that sense is purely data-driven. The ability to represent the clustering process and results through dendrograms is also an advantage of this method, as it is easier to understand and interpret. There are, however, a few limitations to consider:

376 CLUSTER ANALYSIS

1. Hierarchical clustering requires the computation and storage of an n×n distance matrix. For very large datasets, this can be expensive and slow.

2. The hierarchical algorithm makes only one pass through the data. This means that records that are allocated incorrectly early in the process can- not be reallocated subsequently.

3. Hierarchical clustering also tends to have low stability. Reordering data or dropping a few records can lead to a different solution.

4. With respect to the choice of distance between clusters, single and com- plete linkage are robust to changes in the distance metric (e.g., Euclidean, statistical distance) as long as the relative ordering is kept. In contrast, average linkage is more influenced by the choice of distance metric, and might lead to completely different clusters when the metric is changed.

5. Hierarchical clustering is sensitive to outliers.

15.5 Non-hierarchical Clustering: The k-Means Algorithm

A non-hierarchical approach to forming good clusters is to pre-specify a desired number of clusters, k, and assign each case to one of the k clusters so as to minimize a measure of dispersion within the clusters. In other words, the goal is to divide the sample into a predetermined number k of non-overlapping clusters so that clusters are as homogeneous as possible with respect to the measurements used.

A common measure of within-cluster dispersion is the sum of distances (or sum of squared Euclidean distances) of records from their cluster centroid. The problem can be set up as an optimization problem involving integer program- ming, but because solving integer programs with a large number of variables is time-consuming, clusters are often computed using a fast, heuristic method that produces good (although not necessarily optimal) solutions. The k-means algorithm is one such method.

The k-means algorithm starts with an initial partition of the records into k clusters. Subsequent steps modify the partition to reduce the sum of the distances of each record from its cluster centroid. The modification consists of allocating each record to the nearest of the k centroids of the previous partition. This leads to a new partition for which the sum of distances is smaller than before. The means of the new clusters are computed and the improvement step is repeated until the improvement is very small.

NON-HIERARCHICAL CLUSTERING: THE k-MEANS ALGORITHM 377

k- M E A N S C L U S T E R I N G A L G O R I T H M :

1. Start with k initial clusters (user chooses k).

2. At every step, each record is reassigned to the cluster with the “closest” centroid.

3. Recompute the centroids of clusters that lost or gained a record, and repeat Step 2.

4. Stop when moving any more records between clusters increases cluster disper- sion.

Returning to the example with the five utilities and two measurements, let us assume that k = 2 and that the initial clusters are A = {Arizona, Boston} and B = {Central, Commonwealth, Consolidated}. The cluster centroids were computed in Section 15.4:

xA = [−0.516,−0.020] and xB = [−0.733, 0.296]. The distance of each record from each of these two centroids is shown in

Table 15.7.

TABLE 15.7 DISTANCE OF EACH RECORD FROM EACH CENTROID

Distance from Distance from Centroid A Centroid B

Arizona 1.0052 1.3887 Boston 1.0052 0.6216 Central 0.6029 0.8995 Commonwealth 0.7281 1.0207 Consolidated 2.0172 1.6341

We see that Boston is closer to cluster B, and that Central and Common- wealth are each closer to cluster A. We therefore move each of these records to the other cluster and obtain A = {Arizona, Central, Commonwealth} and B = {Consolidated, Boston}. Recalculating the centroids gives

xA = [−0.191,−0.553] and xB = [−1.33, 1.253]. The distance of each record from each of the newly calculated centroids is

given in Table 15.8. At this point we stop, because each record is allocated to its closest cluster.

Choosing the Number of Clusters (k)

The choice of the number of clusters can either be driven by external consider- ations (e.g., previous knowledge, practical constraints, etc.), or we can try a few

378 CLUSTER ANALYSIS

TABLE 15.8 DISTANCE OF EACH RECORD FROM EACH NEWLY CALCULATED CENTROID

Distance from Distance from Centroid A Centroid B

Arizona 0.3827 2.5159 Boston 1.6289 0.5067 Central 0.5463 1.9432 Commonwealth 0.5391 2.0745 Consolidated 2.6412 0.5067

different values for k and compare the resulting clusters. After choosing k, the n records are partitioned into these initial clusters. If there is external reasoning that suggests a certain partitioning, this information should be used. Alterna- tively, if there exists external information on the centroids of the k clusters, this can be used to initially allocate the records.

In many cases, there is no information to be used for the initial partition. In these cases, the algorithm can be rerun with different randomly generated starting partitions to reduce chances of the heuristic producing a poor solution. The number of clusters in the data is generally not known, so it is a good idea to run the algorithm with different values for k that are near the number of clusters that one expects from the data, to see how the sum of distances reduces with increasing values of k. Note that the clusters obtained using different values of k will not be nested (unlike those obtained by hierarchical methods).

The results of running the k-means algorithm using function kmeans() for all 22 utilities and eight measurements with k = 6 are shown in Table 15.9. As in the results from the hierarchical clustering, we see once again that {San Diego} is a singleton cluster. Two more clusters (cluster #3 = {Arizona, Florida, Central, Kentucky, Oklahoma, Texas} and cluster #4 = {Virginia, Northern, Common- wealth, Madison, Wisconsin}) are nearly identical to those that emerged in the hierarchical clustering with average linkage.

To characterize the resulting clusters, we examine the cluster centroids (numerically in Table 15.10 and in the line chart (“profile plot”) in Figure 15.5). We see, for instance, that cluster #5 is characterized by especially low Fixed- charge and RoR, and high Demand-growth. We can also see which variables do the best job of separating the clusters. For example, the spread of clusters for Sales and Fixed-charge is quite high, and not so high for the other variables.

We can also inspect the information on the within-cluster dispersion. From Table 15.10, we see that cluster #3, with seven records, has the largest within- cluster sum of squared distances. In comparison, cluster #6, with three records,

NON-HIERARCHICAL CLUSTERING: THE k-MEANS ALGORITHM 379

TABLE 15.9 k-MEANS CLUSTERING OF 22 UTILITIES INTO k = 6 CLUSTERS (SORTED BY CLUSTER ID

code for k-means

# load and preprocess data utilities.df <- read.csv("Utilities.csv") row.names(utilities.df) <- utilities.df[,1] utilities.df <- utilities.df[,-1]

# normalized distance: utilities.df.norm <- sapply(utilities.df, scale) row.names(utilities.df.norm) <- row.names(utilities.df)

# run kmeans algorithm km <- kmeans(utilities.df.norm, 6)

# show cluster membership km$cluster

Output

> km$cluster Arizona Boston Central Commonwealth NY

3 2 3 4 1 Florida Hawaiian Idaho Kentucky Madison

3 2 6 3 4 Nevada New England Northern Oklahoma Pacific

6 2 4 3 2 Puget San Diego Southern Texas Wisconsin

6 5 3 3 4 United Virginia

2 4

has a smaller within-cluster sum of squared distances. This means that cluster #6 is more homogeneous than cluster #3 (although it is also has fewer records).

When the number of clusters is not predetermined by domain requirements, we can use a graphical approach to evaluate different numbers of clusters. An “elbow chart” is a line chart depicting the decline in cluster heterogeneity as we add more clusters. Figure 15.6 shows the overall average within-cluster distance (normalized) for different choices of k. Moving from 1 to 2 tightens clusters considerably (reflected by the large reduction in within-cluster distance), and so does moving from 2 to 3 and even to 4. Adding more clusters beyond 4 brings less improvement to cluster homogeneity.

From the distances between clusters measured by Euclidean distance between the centroids (see Table 15.11), we can learn about the separation of the different clusters. For instance, we can see that clusters #3 and #4 are the closest to one another, and clusters #1 and #5 (each with a single record) are the most distant from one another. Cluster #5 is most distant from the other clusters, overall, but there is no cluster that is a striking outlier.

380 CLUSTER ANALYSIS

TABLE 15.10 CLUSTER CENTROIDS AND SQUARED DISTANCES FOR k-MEANS WITH k = 6

> # centroids > km$centers

Fixed_charge RoR Cost Load_factor Demand_growth 1 2.03732429 -0.8628882 0.5782326 -1.2950193 -0.7186431 2 -0.35819462 -0.3637904 0.3985832 1.1572643 -0.3017426 3 0.50431607 0.7795509 -0.9858961 -0.3375463 -0.4895769 4 -0.01133215 0.3313815 0.2189339 -0.3580408 0.1664686 5 -1.91907572 -1.9323833 -0.7812761 1.1034665 1.8468982 6 -0.60027572 -0.8331800 1.3389101 -0.4805802 0.9917178

Sales Nuclear Fuel_Cost 1 -1.5814284 0.2143888 1.6926380 2 -0.7080723 -0.4025746 1.1171999 3 0.3518600 -0.5232108 -0.4105368 4 -0.4018738 1.5650384 -0.5954476 5 -0.9014253 -0.2203441 1.4696557 6 1.8565214 -0.7146294 -0.9657660

> # within-cluster sum of squares > km$withinss [1] 0.000000 11.935249 26.507769 10.177094 0.000000 9.533522

> # cluster size > km$size [1] 1 5 7 5 1 3

code for plotting profile plot of centroids

# plot an empty scatter plot plot(c(0), xaxt = 'n', ylab = "", type = "l",

ylim = c(min(km$centers), max(km$centers)), xlim = c(0, 8))

# label x-axes axis(1, at = c(1:8), labels = names(utilities.df))

# plot centroids for (i in c(1:6)) lines(km$centers[i,], lty = i, lwd = 2, col = ifelse(i %in% c(1, 3, 5),

"black", "dark grey"))

# name clusters text(x = 0.5, y = km$centers[, 1], labels = paste("Cluster", c(1:6)))

− 2

− 1

0 1

2

Index

Fixed_charge RoR Cost Load_factor Demand_growth Sales Nuclear Fuel_Cost

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Cluster 5

Cluster 6

FIGURE 15.5 VISUAL PRESENTATION (PROFILE PLOT) OF CLUSTER CENTROIDS

NON-HIERARCHICAL CLUSTERING: THE k-MEANS ALGORITHM 381

TABLE 15.11 EUCLIDEAN DISTANCE BETWEEN CLUSTER CENTROIDS

> dist(km$centers) 1 2 3 4 5

2 3.698906348 3 4.143482338 3.116033813 4 3.983046535 3.170117175 2.704121216 5 5.628591280 3.332157258 5.099264137 4.735628592 6 5.556507647 4.065954404 3.748248485 3.753375354 4.946218349

1 2 3 4 5 6

50 10

0 15

0

Number of Clusters (k)

A ve

ra ge

W ith

in −

C lu

st er

S qu

ar ed

D is

ta nc

e

FIGURE 15.6 COMPARING DIFFERENT CHOICES OF k IN TERMS OF OVERALL AVERAGE WITHIN-CLUSTER DISTANCE

Finally, we can use the information on the distance between the final clusters to evaluate the cluster validity. The ratio of the sum of squared distances for a given k to the sum of squared distances to the mean of all the records (k = 1) is a useful measure for the usefulness of the clustering. If the ratio is near 1.0, the clustering has not been very effective, whereas if it is small, we have well- separated groups.

382 CLUSTER ANALYSIS

PROBLEMS

15.1 University Rankings. The dataset on American College and University Rankings (available from www.dataminingbook.com) contains information on 1302 American colleges and universities offering an undergraduate program. For each university, there are 17 measurements, including continuous measurements (such as tuition and grad- uation rate) and categorical measurements (such as location by state and whether it is a private or public school).

Note that many records are missing some measurements. Our first goal is to estimate these missing values from “similar” records. This will be done by clustering the complete records and then finding the closest cluster for each of the partial records. The missing values will be imputed from the information in that cluster.

a. Remove all records with missing measurements from the dataset.

b. For all the continuous measurements, run hierarchical clustering using complete linkage and Euclidean distance. Make sure to normalize the measurements. From the dendrogram: How many clusters seem reasonable for describing these data?

c. Compare the summary statistics for each cluster and describe each cluster in this context (e.g., “Universities with high tuition, low acceptance rate…”). Hint: To obtain cluster statistics for hierarchical clustering, use the aggregate() function.

d. Use the categorical measurements that were not used in the analysis (State and Pri- vate/Public) to characterize the different clusters. Is there any relationship between the clusters and the categorical information?

e. What other external information can explain the contents of some or all of these clusters?

f. Consider Tufts University, which is missing some information. Compute the Euclidean distance of this record from each of the clusters that you found above (using only the measurements that you have). Which cluster is it closest to? Impute the missing values for Tufts by taking the average of the cluster on those measure- ments.

15.2 Pharmaceutical Industry. An equities analyst is studying the pharmaceutical indus- try and would like your help in exploring and understanding the financial data collected by her firm. Her main objective is to understand the structure of the pharmaceutical industry using some basic financial measures.

Financial data gathered on 21 firms in the pharmaceutical industry are available in the file Pharmaceuticals.csv. For each firm, the following variables are recorded:

1. Market capitalization (in billions of dollars) 2. Beta 3. Price/earnings ratio 4. Return on equity 5. Return on assets 6. Asset turnover 7. Leverage 8. Estimated revenue growth 9. Net profit margin

10. Median recommendation (across major brokerages) 11. Location of firm’s headquarters 12. Stock exchange on which the firm is listed

PROBLEMS 383

Use cluster analysis to explore and analyze the given dataset as follows:

a. Use only the numerical variables (1 to 9) to cluster the 21 firms. Justify the various choices made in conducting the cluster analysis, such as weights for different vari- ables, the specific clustering algorithm(s) used, the number of clusters formed, and so on.

b. Interpret the clusters with respect to the numerical variables used in forming the clusters.

c. Is there a pattern in the clusters with respect to the numerical variables (10 to 12)? (those not used in forming the clusters)

d. Provide an appropriate name for each cluster using any or all of the variables in the dataset.

15.3 Customer Rating of Breakfast Cereals. The datasetCereals.csv includes nutritional information, store display, and consumer ratings for 77 breakfast cereals.

Data Preprocessing. Remove all cereals with missing values.

a. Apply hierarchical clustering to the data using Euclidean distance to the normal- ized measurements. Compare the dendrograms from single linkage and complete linkage, and look at cluster centroids. Comment on the structure of the clusters and on their stability. Hint: To obtain cluster centroids for hierarchical clustering, compute the average values of each cluster members, using the aggregate() function.

b. Which method leads to the most insightful or meaningful clusters?

c. Choose one of the methods. How many clusters would you use? What distance is used for this cutoff? (Look at the dendrogram.)

d. The elementary public schools would like to choose a set of cereals to include in their daily cafeterias. Every day a different cereal is offered, but all cereals should support a healthy diet. For this goal, you are requested to find a cluster of “healthy cereals.” Should the data be normalized? If not, how should they be used in the cluster analysis?

15.4 Marketing to Frequent Fliers. The file EastWestAirlinesCluster.csv contains infor- mation on 3999 passengers who belong to an airline’s frequent flier program. For each passenger, the data include information on their mileage history and on different ways they accrued or spent miles in the last year. The goal is to try to identify clus- ters of passengers that have similar characteristics for the purpose of targeting different segments for different types of mileage offers.

a. Apply hierarchical clustering with Euclidean distance and Ward’s method. Make sure to normalize the data first. How many clusters appear?

b. What would happen if the data were not normalized?

c. Compare the cluster centroid to characterize the different clusters, and try to give each cluster a label.

d. To check the stability of the clusters, remove a random 5% of the data (by taking a random sample of 95% of the records), and repeat the analysis. Does the same picture emerge?

e. Use k-means clustering with the number of clusters that you found above. Does the same picture emerge?

f. Which clusters would you target for offers, and what types of offers would you target to customers in that cluster?

Part VI

Forecasting Time Series

CHAPTER 16

Handling Time Series

In this chapter, we describe the context of business time series forecasting and introduce the main approaches that are detailed in the next chapters, and in par- ticular, regression-based forecasting and smoothing-based methods. Our focus is on forecasting future values of a single time series. These three chapters are meant as an introduction to the general forecasting approach and methods.

In this chapter, we discuss the difference between the predictive nature of time series forecasting vs. the descriptive or explanatory task of time series anal- ysis. A general discussion of combining forecasting methods or results for added precision follows. Next, we present a time series in terms of four components (level, trend, seasonality, and noise) and present methods for visualizing the dif- ferent components and for exploring time series data. We close with a discussion of data partitioning (creating training and validation sets), which is performed differently from cross-sectional data partitioning.

16.1 Introduction1

Time series forecasting is performed in nearly every organization that works with quantifiable data. Retail stores use it to forecast sales. Energy companies use it to forecast reserves, production, demand, and prices. Educational institutions use it to forecast enrollment. Governments use it to forecast tax receipts and spend- ing. International financial organizations like the World Bank and International

1This and subsequent sections in this chapter, copyright © 2017 Datastats, LLC, and Galit Shmueli. Used by permission.

387

388 HANDLING TIME SERIES

Monetary Fund use it to forecast inflation and economic activity. Transportation companies use time series forecasting to forecast future travel. Banks and lend- ing institutions use it (sometimes badly!) to forecast new home purchases. And venture capital firms use it to forecast market potential and to evaluate business plans.

Previous chapters in this book deal with classifying and predicting data where time is not a factor, in the sense that it is not treated differently from other variables, and where the sequence of measurements over time does not matter. These are typically called cross-sectional data. In contrast, this chapter deals with a different type of data: time series.

With today’s technology, many time series are recorded on very frequent time scales. Stock data are available at ticker level. Purchases at online and offline stores are recorded in real time. The Internet of Things (IoT) gener- ates huge numbers of time series, produced by sensors and other measurements devices. Although data might be available at a very frequent scale, for the pur- pose of forecasting, it is not always preferable to use this time scale. In con- sidering the choice of time scale, one must consider the scale of the required forecasts and the level of noise in the data. For example, if the goal is to fore- cast next-day sales at a grocery store, using minute-by-minute sales data is likely to be less useful for forecasting than using daily aggregates. The minute-by- minute series will contain many sources of noise (e.g., variation by peak and non-peak shopping hours) that degrade its forecasting power, and these noise errors, when the data are aggregated to a cruder level, are likely to average out.

The focus in this part of the book is on forecasting a single time series. In some cases, multiple time series are to be forecasted (e.g., the monthly sales of multiple products). Even when multiple series are being forecasted, the most popular forecasting practice is to forecast each series individually. The advan- tage of single-series forecasting is its simplicity. The disadvantage is that it does not take into account possible relationships between series. The statistics litera- ture contains models for multivariate time series which directly model the cross- correlations between series. Such methods tend to make restrictive assumptions about the data and the cross-series structure. They also require statistical expertise for estimation and maintenance. Econometric models often include information from one or more series as inputs into another series. However, such models are based on assumptions of causality that are based on theoretical models. An alternative approach is to capture the associations between the series of interest and external information more heuristically. An example is using the sales of lipstick to forecast some measure of the economy, based on the observation by Ronald Lauder, chairman of Estee Lauder, that lipstick sales tend to increase before tough economic times (a phenomenon called the “leading lipstick indicator”).

DESCRIPTIVE VS. PREDICTIVE MODELING 389

16.2 Descriptive vs. Predictive Modeling

As with cross-sectional data, modeling time series data is done for either descrip- tive or predictive purposes. In descriptive modeling, or time series analysis, a time series is modeled to determine its components in terms of seasonal patterns, trends, relation to external factors, etc. These can then be used for decision- making and policy formulation. In contrast, time series forecasting uses the infor- mation in a time series (and perhaps other information) to forecast future values of that series. The difference between the goals of time series analysis and time series forecasting leads to differences in the type of methods used and in the modeling process itself. For example, in selecting a method for describing a time series, priority is given to methods that produce understandable results (rather than “blackbox” methods) and sometimes to models based on causal arguments (explanatory models). Furthermore, describing can be done in retrospect, while forecasting is prospective in nature. This means that descriptive models might use “future” information (e.g., averaging the values of yesterday, today, and tomor- row to obtain a smooth representation of today’s value) whereas forecasting mod- els cannot.

The focus in this chapter is on time series forecasting, where the goal is to predict future values of a time series. For information on time series analysis, see Chatfield (2003).

16.3 Popular Forecasting Methods in Business

In this part of the book, we focus on two main types of forecasting methods that are popular in business applications. Both are versatile and powerful, yet rela- tively simple to understand and deploy. One type of forecasting tool is multiple linear regression, where the user specifies a certain model and then estimates it from the time series. The other is the more data-driven tool of smoothing, where the method learns patterns from the data. Each of the two types of tools has advantages and disadvantages, as detailed in Chapters 17 and 18. We also note that data mining methods such as neural networks and others that are intended for cross-sectional data are also sometimes used for time series forecasting, espe- cially for incorporating external information into the forecasts (see Shmueli & Lichtendahl, 2016).

Combining Methods

Before a discussion of specific forecasting methods in the following two chap- ters, it should be noted that a popular approach for improving predictive per- formance is to combine forecasting methods. This is similar to the ensembles

390 HANDLING TIME SERIES

approach described in Chapter 13. Combining forecasting methods can be done via two-level (or multi level) forecasters, where the first method uses the origi- nal time series to generate forecasts of future values, and the second method uses the residuals from the first model to generate forecasts of future forecast errors, thereby “correcting” the first level forecasts. We describe two-level forecasting in Chapter 17 (Section 17.4). Another combination approach is via “ensem- bles,” where multiple methods are applied to the time series, and their resulting forecasts are averaged in some way to produce the final forecast. Combining methods can take advantage of the strengths of different forecasting methods to capture different aspects of the time series (also true in cross-sectional data). The averaging across multiple methods can lead to forecasts that are more robust and of higher precision.

16.4 Time Series Components

In both types of forecasting methods, regression models and smoothing, and in general, it is customary to dissect a time series into four components: level, trend, seasonality, and noise. The first three components are assumed to be invisible, as they characterize the underlying series, which we only observe with added noise. Level describes the average value of the series, trend is the change in the series from one period to the next, and seasonality describes a short-term cyclical behavior of the series which can be observed several times within the given series. Finally, noise is the random variation that results from measurement error or other causes not accounted for. It is always present in a time series to some degree.

Don’t get confused by the standard conversational meaning of ”season” (winter, spring, etc.). The statistical meaning of ”season” refers to any time period that repeats itself in cycles within the larger time series. Also, note that the term ”period” in forecasting means simply a span of time, and not the specific meaning that it has in physics, of the distance between two equivalent points in a cycle.

In order to identify the components of a time series, the first step is to examine a time plot. In its simplest form, a time plot is a line chart of the series values over time, with temporal labels (e.g., calendar date) on the horizontal axis. To illustrate this, consider the following example.

Example: Ridership on Amtrak Trains

Amtrak, a US railway company, routinely collects data on ridership. Here we focus on forecasting future ridership using the series of monthly ridership

TIME SERIES COMPONENTS 391

between January 1991 and March 2004. These data are publicly available at www.forecastingprinciples.com.2

A time plot for the monthly Amtrak ridership series is shown in Figure 16.1. Note that the values are in thousands of riders.

code for creating a time series plot

library(forecast) Amtrak.data <- read.csv("Amtrak.csv")

# create time series object using ts() # ts() takes three arguments: start, end, and freq. # with monthly data, the frequency of periods per season is 12 (per year). # arguments start and end are (season number, period number) pairs. # here start is Jan 1991: start = c(1991, 1); end is Mar 2004: end = c(2004, 3). ridership.ts <- ts(Amtrak.data$Ridership,

start = c(1991, 1), end = c(2004, 3), freq = 12)

# plot the series plot(ridership.ts, xlab = "Time", ylab = "Ridership (in 000s)", ylim = c(1300, 2300))

Time

R id

er sh

ip (

in 0

00 s)

1992 1994 1996 1998 2000 2002 2004

14 00

16 00

18 00

20 00

22 00

FIGURE 16.1 MONTHLY RIDERSHIP ON AMTRAK TRAINS (IN THOUSANDS) FROM JANUARY 1991 TO MARCH 2004

2To get this series: click on Data. Under T-Competition Data, click “time-series data” (the direct URL to the data file is www.forecastingprinciples.com/files/MHcomp1.xls as of July 2015). This file contains many time series. In the Monthly worksheet, column AI contains series M034.

392 HANDLING TIME SERIES

Looking at the time plot reveals the nature of the series components: the overall level is around 1,800,000 passengers per month. A slight U-shaped trend is discernible during this period, with pronounced annual seasonality, with peak travel during summer (July and August).

A second step in visualizing a time series is to examine it more carefully. A few tools are useful:

Zoom in: Zooming in to a shorter period within the series can reveal patterns that are hidden when viewing the entire series. This is especially important when the time series is long. Consider a series of the daily number of vehicles passing through the Baregg tunnel in Switzerland (data are available in the same location as the Amtrak Ridership data; series D028). The series from November 1, 2003 to November 16, 2005 is shown in the top panel of Figure 16.2. Zooming in to a 4-month period (bottom panel) reveals a strong day-of-week pattern that is not visible in the time plot of the complete time series.

60 00

0 10

00 00

14 00

00

Day

N um

be r

of V

eh ic

le s

01−Nov−03 01−Jul−04 01−Mar−05 01−Nov−05

60 00

0 10

00 00

14 00

00

Day

N um

be r

of V

eh ic

le s

01−Feb−04 01−Mar−04 01−Apr−04 01−May−04 31−May−04

FIGURE 16.2 TIME PLOTS OF THE DAILY NUMBER OF VEHICLES PASSING THROUGH THE BAREGG TUNNEL, SWITZERLAND. THE BOTTOM PANEL ZOOMS IN TO A 4-MONTH PERIOD, REVEALING A DAY-OF-WEEK PATTERN

TIME SERIES COMPONENTS 393

Change scale of series: In order to better identify the shape of a trend, it is useful to change the scale of the series. One simple option is to change the vertical scale to an exponential scale (in Excel Ribbon, select Chart Tools > Axes > Primary Vertical Axis > More Primary Vertical Axis Options, and click “logarithmic scale” in the Format Axis menu). If the trend on the new scale appears more linear, then the trend in the original series is closer to an exponential trend.

Add trend lines: Another possibility for better discerning the shape of the trend is to add a trend line (Excel Ribbon: Chart Tools > Trendline). By trying different trendlines, one can see what type of trend (e.g., linear, expo- nential, quadratic) best approximates the data.

Suppress seasonality: It is often easier to see trends in the data when seasonality is suppressed. Suppressing seasonality patterns can be done by plotting the series at a cruder time scale (e.g., aggregating monthly data into years) or creating separate lines or time plots for each season (e.g., separate lines for each day of week). Another popular option is to use moving average charts. We will discuss these in Chapter 18 (Section 18.2).

Continuing our example of Amtrak ridership, the charts in Figure 16.3 help make the series’ components more visible.

Some forecasting methods directly model these components by making assumptions about their structure. For example, a popular assumption about trend is that it is linear or exponential over the given time period or part of it. Another common assumption is about the noise structure: many statisti- cal methods assume that the noise follows a normal distribution. The advan- tage of methods that rely on such assumptions is that when the assumptions are reasonably met, the resulting forecasts will be more robust and the mod- els more understandable. Other forecasting methods, which are data-adaptive, make fewer assumptions about the structure of these components, and instead try to estimate them only from the data. Data-adaptive methods are advanta- geous when such assumptions are likely to be violated, or when the structure of the time series changes over time. Another advantage of many data-adaptive methods is their simplicity and computational efficiency.

A key criterion for deciding between model-driven and data-driven fore- casting methods is the nature of the series in terms of global vs. local patterns. A global pattern is one that is relatively constant throughout the series. An example is a linear trend throughout the entire series. In contrast, a local pattern is one that occurs only in a short period of the data, and then changes, for example, a trend that is approximately linear within four neighboring time points, but the trend size (slope) changes slowly over time.

Model-driven methods are generally preferable for forecasting series with global patterns as they use all the data to estimate the global pattern. For a local

394 HANDLING TIME SERIES

code for creating Figure 16.3

library(forecast)

# create short time series # use window() to create a new, shorter time series of ridership.ts # for the new three-year series, start time is Jan 1997 and end time is Dec 1999 ridership.ts.3yrs <- window(ridership.ts, start = c(1997, 1), end = c(1999, 12))

# fit a linear regression model to the time series ridership.lm <- tslm(ridership.ts ~ trend + I(trend^2))

# shorter and longer time series par(mfrow = c(2, 1)) plot(ridership.ts.3yrs, xlab = "Time", ylab = "Ridership (in 000s)",

ylim = c(1300, 2300)) plot(ridership.ts, xlab = "Time", ylab = "Ridership (in 000s)", ylim = c(1300, 2300))

# overlay the fitted values of the linear model lines(ridership.lm$fitted, lwd = 2)

Time

R id

er sh

ip (

in 0

00 s)

1997.0 1997.5 1998.0 1998.5 1999.0 1999.5 2000.0

14 00

18 00

22 00

Time

R id

er sh

ip (

in 0

00 s)

1992 1994 1996 1998 2000 2002 2004

14 00

18 00

22 00

FIGURE 16.3 PLOTS THAT ENHANCE THE DIFFERENT COMPONENTS OF THE TIME SERIES. TOP: ZOOM-IN TO 3 YEARS OF DATA. BOTTOM: ORIGINAL SERIES WITH OVERLAID QUADRATIC TRENDLINE

DATA-PARTITIONING AND PERFORMANCE EVALUATION 395

pattern, a model-driven model would require specifying how and when the patterns change, which is usually impractical and often unknown. Therefore, data-driven methods are preferable for local patterns. Such methods “learn” patterns from the data and their memory length can be set to best adapt to the rate of change in the series. Patterns that change quickly warrant a “short memory,” whereas patterns that change slowly warrant a “long memory.” In conclusion, the time plot should be used not only to identify the time series component, but also the global/local nature of the trend and seasonality.

16.5 Data-Partitioning and Performance Evaluation

As in the case of cross-sectional data, in order to avoid overfitting and to be able to assess the predictive performance of the model on new data, we first partition the data into a training set and a validation set (and perhaps an additional test set). However, there is one important difference between data partitioning in cross-sectional and time series data. In cross-sectional data, the partitioning is usually done randomly, with a random set of records designated as training data and the remainder as validation data. However, in time series, a random parti- tion would create two time series with “holes.” Nearly all standard forecasting methods cannot handle time series with missing values. Therefore, we partition a time series into training and validation sets differently. The series is trimmed into two periods; the earlier period is set as the training data and the later period as the validation data. Methods are then trained on the earlier training period, and their predictive performance assessed on the later validation period. Evalu- ation measures typically use the same metrics used in cross-sectional evaluation (see Chapter 5) with MAE, MAPE, and RMSE being the most popular metrics in practice. In evaluating and comparing forecasting methods, another impor- tant tool is visualization: examining time plots of the actual and predicted series can shed light on performance and hint toward possible improvements.

Benchmark Performance: Naive Forecasts

While it is tempting to apply “sophisticated” forecasting methods, we must eval- uate their value added compared to a very simple approach: the naive forecast. A naive forecast is simply the most recent value of the series. In other words, at time t, our forecast for any future period t+k is simply the value of the series at time t. While simple, naive forecasts are sometimes surprisingly difficult to out- perform with more sophisticated models. It is therefore important to benchmark against results from a naive-forecasting approach.

When a time series has seasonality, a seasonal naive forecast can be generated. It is simply the last similar value in the season. For example, to forecast April

396 HANDLING TIME SERIES

2001 for Amtrak ridership, we use the ridership from the most recent April, April 2000. Similarly, to forecast April 2002, we also use April 2000 ridership. In Figure 16.4, we show naive (horizontal line) and seasonal naive forecasts, as well as actual values (dotted line), in a 3-year validation set from April 2001 to March 2004.

Table 16.1 compares the accuracies of these two naive forecasts. Because Amtrak ridership has monthly seasonality, the seasonal naive forecast is the clear winner on both training and validation sets and on all popular measures. In choosing between the two models, the accuracy on the validation set is more relevant than the accuracy on the training set. Performance on the validation set is more indicative of how the models will perform in the future.

TABLE 16.1 PREDICTIVE ACCURACY OF NAIVE AND SEASONAL NAIVE FORECASTS IN THE VALIDATION SET FOR AMTRAK RIDERSHIP

> accuracy(naive.pred, valid.ts) ME RMSE MAE MPE MAPE

Training set 2.45091 168.1470 125.2975 -0.3460027 7.271393 Test set -14.71772 142.7551 115.9234 -1.2749992 6.021396 > accuracy(snaive.pred, valid.ts)

ME RMSE MAE MPE MAPE Training set 13.93991 99.26557 82.49196 0.5850656 4.715251 Test set 54.72961 95.62433 84.09406 2.6527928 4.247656

Generating Future Forecasts

Another important difference between cross-sectional and time-series partition- ing occurs when creating the actual forecasts. Before attempting to forecast future values of the series, the training and validation sets are recombined into one long series, and the chosen method/model is rerun on the complete data. This final model is then used to forecast future values. The three advantages in recombining are:

1. The validation set, which is the most recent period, usually contains the most valuable information in terms of being the closest in time to the forecast period;

2. With more data (the complete time series compared to only the training set), some models can be estimated more accurately;

3. If only the training set is used to generate forecasts, then it will require forecasting farther into the future (e.g., if the validation set contains four time points, forecasting the next period will require a 5-step-ahead fore- cast from the training set).

DATA-PARTITIONING AND PERFORMANCE EVALUATION 397

code for creating Figure 16.4

nValid <- 36 nTrain <- length(ridership.ts) - nValid

# partition the data train.ts <- window(ridership.ts, start = c(1991, 1), end = c(1991, nTrain)) valid.ts <- window(ridership.ts, start = c(1991, nTrain + 1),

end = c(1991, nTrain + nValid))

# generate the naive and seasonal naive forecasts naive.pred <- naive(train.ts, h = nValid) snaive.pred <- snaive(train.ts, h = nValid)

# plot forecasts and actuals in the training and validation sets plot(train.ts, ylim = c(1300, 2600), ylab = "Ridership", xlab = "Time", bty = "l",

xaxt = "n", xlim = c(1991,2006.25), main = "") axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(naive.pred$mean, lwd = 2, col = "blue", lty = 1) lines(snaive.pred$mean, lwd = 2, col = "blue", lty = 1) lines(valid.ts, col = "grey20", lty = 3) lines(c(2004.25 - 3, 2004.25 - 3), c(0, 3500)) lines(c(2004.25, 2004.25), c(0, 3500)) text(1996.25, 2500, "Training") text(2002.75, 2500, "Validation") text(2005.25, 2500, "Future") arrows(2004 - 3, 2450, 1991.25, 2450, code = 3, length = 0.1, lwd = 1,angle = 30) arrows(2004.5 - 3, 2450, 2004, 2450, code = 3, length = 0.1, lwd = 1,angle = 30) arrows(2004.5, 2450, 2006, 2450, code = 3, length = 0.1, lwd = 1, angle = 30)

Time

R id

er sh

ip

14 00

16 00

18 00

20 00

22 00

24 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 16.4 NAIVE AND SEASONAL NAIVE FORECASTS IN A 3-YEAR VALIDATION SET FOR AMTRAK RIDERSHIP

398 HANDLING TIME SERIES

PROBLEMS

16.1 Impact of September 11 on Air Travel in the United States. The Research and Innovative Technology Administration’s Bureau of Transportation Statistics conducted a study to evaluate the impact of the September 11, 2001 terrorist attack on US trans- portation. The 2006 study report and the data can be found at http://goo.gl/w2lJPV. The goal of the study was stated as follows:

The purpose of this study is to provide a greater understanding of the passenger travel behavior patterns of persons making long distance trips before and after 9/11.

The report analyzes monthly passenger movement data between January 1990 and May 2004. Data on three monthly time series are given in file Sept11Travel.csv for this period:

(1) Actual airline revenue passenger miles (Air),

(2) Rail passenger miles (Rail), and

(3) Vehicle miles traveled (Car).

In order to assess the impact of September 11, BTS took the following approach: using data before September 11, they forecasted future data (under the assumption of no terrorist attack). Then, they compared the forecasted series with the actual data to assess the impact of the event. Our first step, therefore, is to split each of the time series into two parts: pre- and post September 11. We now concentrate only on the earlier time series.

a. Is the goal of this study descriptive or predictive?

b. Plot each of the three pre-event time series (Air, Rail, Car).

i. What time series components appear from the plot?

ii. What type of trend appears? Change the scale of the series, add trendlines and suppress seasonality to better visualize the trend pattern.

16.2 Performance on Training and Validation Data. Two different models were fit to the same time series. The first 100 time periods were used for the training set and the last 12 periods were treated as a hold-out set. Assume that both models make sense practically and fit the data pretty well. Below are the RMSE values for each of the models:

Training Set Validation Set

Model A 543 690 Model B 669 675

a. Which model appears more useful for explaining the different components of this time series? Why?

b. Which model appears to be more useful for forecasting purposes? Why?

16.3 Forecasting Department Store Sales. The file DepartmentStoreSales.csv contains data on the quarterly sales for a department store over a 6-year period (data courtesy of Chris Albright).

PROBLEMS 399

a. Create a well-formatted time plot of the data.

b. Which of the four components (level, trend, seasonality, noise) seem to be present in this series?

16.4 Shipments of Household Appliances. The file ApplianceShipments.csv contains the series of quarterly shipments (in million $) of US household appliances between 1985 and 1989 (data courtesy of Ken Black).

a. Create a well-formatted time plot of the data.

b. Which of the four components (level, trend, seasonality, noise) seem to be present in this series?

16.5 Canadian Manufacturing Workers Workhours. The time plot in Figure 16.5 describes the average annual number of weekly hours spent by Canadian manufactur- ing workers (data are available in CanadianWorkHours.csv—thanks to Ken Black for the data).

Year

H ou

rs P

er W

ee k

1970 1975 1980 1985 1990 1995 2000

34 .5

35 .0

35 .5

36 .0

36 .5

37 .0

37 .5

38 . 0

FIGURE 16.5 AVERAGE ANNUAL WEEKLY HOURS SPENT BY CANADIAN MANUFACTURING WORKERS

a. Reproduce the time plot.

b. Which of the four components (level, trend, seasonality, noise) seem to be present in this series?

16.6 Souvenir Sales: The file SouvenirSales.csv contains monthly sales for a souvenir shop at a beach resort town in Queensland, Australia, between 1995–2001. (Source: Hyndman, R.J., Time Series Data Library, http://data.is/TSDLdemo. Accessed on 07/25/15.)

Back in 2001, the store wanted to use the data to forecast sales for the next 12 months (year 2002). They hired an analyst to generate forecasts. The analyst first partitioned the data into training and validation sets, with the validation set containing

400 HANDLING TIME SERIES

the last 12 months of data (year 2001). She then fit a regression model to sales, using the training set.

a. Create a well-formatted time plot of the data.

b. Change the scale on the x-axis, or on the y-axis, or on both to log-scale in order to achieve a linear relationship. Select the time plot that seems most linear.

c. Comparing the two time plots, what can be said about the type of trend in the data?

d. Why were the data partitioned? Partition the data into the training and validation set as explained above.

16.7 Shampoo Sales. The file ShampooSales.csv contains data on the monthly sales of a certain shampoo over a 3-year period. Source: Hyndman, R.J., Time Series Data Library, http://data.is/TSDLdemo. Accessed on 07/25/15).

a. Create a well-formatted time plot of the data.

c. Do you expect to see seasonality in sales of shampoo? Why?

d. If the goal is forecasting sales in future months, which of the following steps should be taken?

• Partition the data into training and validation sets • Tweak the model parameters to obtain good fit to the validation data • Look at MAPE and RMSE values for the training set • Look at MAPE and RMSE values for the validation set

CHAPTER 17

Regression-Based Forecasting

A popular forecasting tool is based on multiple linear regression models, using suitable predictors to capture trend and/or seasonality. In this chapter, we show how a linear regression model can be set up to capture a time series with a trend and/or seasonality. The model, which is estimated from the data, can then pro- duce future forecasts by inserting the relevant predictor information into the estimated regression equation. We describe different types of common trends (linear, exponential, polynomial), as well as two types of seasonality (additive and multiplicative). Next, we show how a regression model can be used to quantify the correlation between neighboring values in a time series (called autocorrela- tion). This type of model, called an autoregressive model, is useful for improving forecast precision by making use of the information contained in the autocor- relation (beyond trend and seasonality). It is also useful for evaluating the pre- dictability of a series, by evaluating whether the series is a “random walk.” The various steps of fitting linear regression and autoregressive models, using them to generate forecasts, and assessing their predictive accuracy, are illustrated using the Amtrak ridership series.

17.1 A Model with Trend1

Linear Trend

To create a linear regression model that captures a time series with a global linear trend, the outcome variable (Y ) is set as the time series values or some function of it, and the predictor (X) is set as a time index. Let us consider a simple

1This and subsequent sections in this chapter copyright © 2017 Datastats, LLC, and Galit Shmueli. Used by permission.

401

402 REGRESSION-BASED FORECASTING

code for creating Figure 17.1

library(forecast) Amtrak.data <- read.csv("Amtrak.csv")

# create time series ridership.ts <- ts(Amtrak.data$Ridership, start = c(1991,1),

end = c(2004,3), freq = 12)

# produce linear trend model ridership.lm <- tslm(ridership.ts ~ trend)

# plot the series plot(ridership.ts, xlab = "Time", ylab = "Ridership", ylim = c(1300,2300),

bty = "l") lines(ridership.lm$fitted, lwd = 2)

Time

R id

er sh

ip

1992 1994 1996 1998 2000 2002 2004

14 00

16 00

18 00

20 00

22 00

FIGURE 17.1 A LINEAR TREND FIT TO AMTRAK RIDERSHIP

example: fitting a linear trend to the Amtrak ridership data. This type of trend is shown in Figure 17.1.

From the time plot, it is obvious that the global trend is not linear. However, we use this example to illustrate how a linear trend is fitted, and later we consider more appropriate models for this series.

A MODEL WITH TREND 403

To obtain a linear relationship between Ridership and Time, we set the output variable Y as the Amtrak Ridership and create a new variable that is a time index t = 1, 2, 3, . . . This time index is then used as a single predictor in the regression model:

Yt = β0 + β1t+ ϵ,

where Yt is the Ridership at period t and ϵ is the standard noise term in a linear regression. Thus, we are modeling three of the four time series components: level (β0), trend (β1), and noise (ϵ). Seasonality is not modeled. A snapshot of the two corresponding columns (Y and t) are shown in Table 17.1.

TABLE 17.1 OUTCOME VARIABLE (MIDDLE) AND PREDICTOR VARIABLE (RIGHT) USED TO FIT A LINEAR TREND

Month Ridership (Yt) t

Jan 91 1709 1 Feb 91 1621 2 Mar 91 1973 3 Apr 91 1812 4 May 91 1975 5 Jun 91 1862 6 Jul 91 1940 7 Aug 91 2013 8 Sep 91 1596 9 Oct 91 1725 10 Nov 91 1676 11 Dec 91 1814 12 Jan 92 1615 13 Feb 92 1557 14

After partitioning the data into training and validation sets, the next step is to fit a linear regression model to the training set, with t as the single predictor (function tslm() relies on ts() which automatically creates t and calls it trend). Applying this to the Amtrak ridership data (with a validation set consisting of the last 12 months) results in the estimated model shown in Figure 17.2. The actual and fitted values and the residuals (or forecast errors) are shown in the two time plots.

Table 17.2 contains a report of the estimated coefficients. Note that exam- ining only the estimated coefficients and their statistical significance can be misleading! In this example, they would indicate that the linear fit is reason- able, although it is obvious from the time plots that the trend is not linear. An inadequate trend shape is easiest to detect by examining the time plot of the residuals.

404 REGRESSION-BASED FORECASTING

code for creating Figure 17.2

# fit linear trend model to training set and create forecasts train.lm <- tslm(train.ts ~ trend) train.lm.pred <- forecast(train.lm, h = nValid, level = 0)

par(mfrow = c(2, 1)) plot(train.lm.pred, ylim = c(1300, 2600), ylab = "Ridership", xlab = "Time",

bty = "l", xaxt = "n", xlim = c(1991,2006.25), main = "", flty = 2) axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(train.lm.pred$fitted, lwd = 2, col = "blue") lines(valid.ts) plot(train.lm.pred$residuals, ylim = c(-420, 500), ylab = "Forecast Errors",

xlab = "Time", bty = "l", xaxt = "n", xlim = c(1991,2006.25), main = "") axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(valid.ts - train.lm.pred$mean, lwd = 1)

Code for data partition is given in Figure 16.4

Time

R id

er sh

ip

14 00

20 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

Time

F or

ec as

t E rr

or s

− 40

0 0

40 0

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 17.2 A LINEAR TREND FIT TO AMTRAK RIDERSHIP IN THE TRAINING PERIOD AND FORECASTED IN THE VALIDATION PERIOD

A MODEL WITH TREND 405

TABLE 17.2 SUMMARY OF OUTPUT FROM A LINEAR REGRESSION MODEL APPLIED TO THE AMTRAK RIDERSHIP DATA IN THE TRAINING PERIOD

> summary(train.lm)

Call: lm(formula = formula, data = "train.ts", na.action = na.exclude)

Residuals: Min 1Q Median 3Q Max

-411.29 -114.02 16.06 129.28 306.35

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) 1750.3595 29.0729 60.206 <2e-16 *** trend 0.3514 0.4069 0.864 0.39 --- Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 160.2 on 121 degrees of freedom Multiple R-squared: 0.006125, Adjusted R-squared: -0.002089 F-statistic: 0.7456 on 1 and 121 DF, p-value: 0.3896

L I N E A R R E G R E S S I O N F O R T I M E S E R I E S I N R

Fitting a regression model to time series (and using it to generate forecasts) can be done in R using the function tslm() in the forecast package.

Exponential Trend

Several alternative trend shapes are useful and easy to fit via a linear regression model. One such shape is an exponential trend. An exponential trend implies a multiplicative increase/decrease of the series over time (Yt = ceβ1t+ϵ). To fit an exponential trend, simply replace the outcome variable Y with logY (where log is the natural logarithm), and fit a linear regression (logYt = β0+β1t+ϵ). In the Amtrak example, for instance, we would fit a linear regression of log(Ridership) on the index variable t. Exponential trends are popular in sales data, where they reflect percentage growth. In R, fitting exponential trend is done by setting argument lambda = 0 in function tslm().2

Note: As in the general case of linear regression, when comparing the pre- dictive accuracy of models that have a different output variable, such as a linear

2Argument lambda is used to apply the Box-Cox transformation to the values of the time series: (yλ− 1)/λ if λ ≠ 0. When λ = 0, the transformation is defined as log(y). When λ = 1, the series is not transformed (except for the subtraction of 1 from each value), so the model has a linear trend.

406 REGRESSION-BASED FORECASTING

trend model (with Y ) and an exponential trend model (with logY ), it is essential to compare forecasts or forecast errors on the same scale. In R, when using an exponential trend model, the forecasts of logY are made and then automatically converted back to the original scale. An example is shown in Figure 17.3, where an exponential trend is fit to the Amtrak ridership data.

code for creating Figure 17.3

# fit exponential trend using tslm() with argument lambda = 0 train.lm.expo.trend <- tslm(train.ts ~ trend, lambda = 0) train.lm.expo.trend.pred <- forecast(train.lm.expo.trend, h = nValid, level = 0)

# fit linear trend using tslm() with argument lambda = 1 (no transform of y) train.lm.linear.trend <- tslm(train.ts ~ trend, lambda = 1) train.lm.linear.trend.pred <- forecast(train.lm.linear.trend, h = nValid, level = 0)

plot(train.lm.expo.trend.pred, ylim = c(1300, 2600), ylab = "Ridership", xlab = "Time", bty = "l", xaxt = "n", xlim = c(1991,2006.25), main = "", flty = 2) axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(train.lm.expo.trend.pred$fitted, lwd = 2, col = "blue") # Added in 6-5 lines(train.lm.linear.trend.pred$fitted, lwd = 2, col = "black", lty = 3) lines(train.lm.linear.trend.pred$mean, lwd = 2, col = "black", lty = 3) lines(valid.ts)

Time

R id

er sh

ip

14 00

16 00

18 00

20 00

22 00

24 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 17.3 EXPONENTIAL (AND LINEAR) TREND USED TO FORECAST AMTRAK RIDERSHIP

A MODEL WITH SEASONALITY 407

Polynomial Trend

Another non-linear trend shape that is easy to fit via linear regression is a poly- nomial trend, and in particular, a quadratic relationship of the form Yt = β0 + β1t + β2t

2 + ϵ. This is done by creating an additional predictor t2 (the square of t), and fitting a multiple linear regression with the two predictors t and t2. In R, we fit a quadratic trend using function I(), which treats an object “as is” (Figure 17.4). For the Amtrak ridership data, we have already seen a U-shaped trend in the data. We therefore fit a quadratic model. The fitted and residual charts are shown in Figure 17.4. We conclude from these plots that this shape adequately captures the trend. The forecast errors are now devoid of trend and exhibit only seasonality.

In general, any type of trend shape can be fit as long as it has a mathematical representation. However, the underlying assumption is that this shape is appli- cable throughout the period of data that we have and also during the period that we are going to forecast. Do not choose an overly complex shape. Although it will fit the training data well, it will in fact be overfitting them. To avoid over- fitting, always examine the validation performance and refrain from choosing overly complex trend patterns.

17.2 A Model with Seasonality

A seasonal pattern in a time series means that observations that fall in some seasons have consistently higher or lower values than those that fall in other sea- sons. Examples are day-of-week patterns, monthly patterns, and quarterly pat- terns. The Amtrak ridership monthly time series, as can be seen in the time plot, exhibits strong monthly seasonality (with highest traffic during summer months).

Seasonality is captured in a regression model by creating a new categorical variable that denotes the season for each value. This categorical variable is then turned into dummies, which in turn are included as predictors in the regression model. To illustrate this, we created a new “Season” column for the Amtrak ridership data, as shown in Table 17.3. Then, to include the Season categorical variable as a predictor in a regression model for Y (Ridership), we turn it into dummies (for m = 12 seasons we create 11 dummies, which are binary vari- ables that take on the value 1 if the record falls in that particular season, and 0 otherwise3).

In R, function tslm() uses ts() which automatically creates the categorical Season column (called season) and converts it into dummy variables.

3We use only m-1 dummies because information about the m − 1 seasons is sufficient. If all m − 1 variables are zero, then the season must be themth season. Including themth dummy causes redundant information and multicollinearity.

408 REGRESSION-BASED FORECASTING

code for creating Figure 17.4

# fit quadratic trend using function I(), which treats an object "as is". train.lm.poly.trend <- tslm(train.ts ~ trend + I(trend^2)) summary(train.lm.poly.trend) train.lm.poly.trend.pred <- forecast(train.lm.poly.trend, h = nValid, level = 0)

par(mfrow = c(2,1)) plot(train.lm.poly.trend.pred, ylim = c(1300, 2600), ylab = "Ridership", xlab = "Time", bty = "l", xaxt = "n", xlim = c(1991,2006.25), main = "", flty = 2) axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(train.lm.poly.trend$fitted, lwd = 2) lines(valid.ts)

plot(train.lm.poly.trend$residuals, ylim = c(-400, 550), ylab = "Forecast Errors", xlab = "Time", bty = "l", xaxt = "n", xlim = c(1991,2006.25), main = "") axis(1, at = seq(1991, 2006, 1), labels = format(seq(1991, 2006, 1))) lines(valid.ts - train.lm.poly.trend.pred$mean, lwd = 1)

Time

R id

er sh

ip

14 00

20 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

Time

F or

ec as

t E rr

or s

− 40

0 0

40 0

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 17.4 QUADRATIC TREND MODEL USED TO FORECAST AMTRAK RIDERSHIP. PLOTS OF FITTED, FORECASTED, AND ACTUAL VALUES (TOP) AND FORECAST ERRORS (BOTTOM)

A MODEL WITH SEASONALITY 409

TABLE 17.3 NEW CATEGORICAL VARIABLE (RIGHT) TO BE USED (VIA DUMMIES) AS PREDICTOR(S) IN A LINEAR REGRESSION MODEL

Month Ridership Season

Jan 91 1709 Jan Feb 91 1621 Feb Mar 91 1973 Mar Apr 91 1812 Apr May 91 1975 May Jun 91 1862 Jun Jul 91 1940 Jul Aug 91 2013 Aug Sep 91 1596 Sep Oct 91 1725 Oct Nov 91 1676 Nov Dec 91 1814 Dec Jan 92 1615 Jan Feb 92 1557 Feb Mar 92 1891 Mar Apr 92 1956 Apr May 92 1885 May

Time

R id

er sh

ip

14 00

20 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

Time

F or

ec as

t E rr

or s

− 40

0 0

40 0

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 17.5 REGRESSION MODEL WITH SEASONALITY APPLIED TO THE AMTRAK RIDERSHIP (TOP) AND ITS FORECAST ERRORS (BOTTOM)

410 REGRESSION-BASED FORECASTING

After partitioning the data into training and validation sets (see Section 16.5), we fit the regression model to the training data. The fitted series and the residuals from this model are shown in Figure 17.5. The model appears to capture the seasonality in the data. However, since we have not included a trend component in the model (as shown in Section 17.1), the fitted values do not capture the existing trend. Therefore, the residuals, which are the difference between the actual and the fitted values, clearly display the remaining U-shaped trend.

When seasonality is added as described above (create categorical seasonal variable, then create dummies from it, then regress on Y ), it captures additive seasonality. This means that the average value of Y in a certain season is a fixed amount more or less than that in another season. Table 17.4 shows the output of a linear regression fit to Ridership (Y ) with seasonality. For example, in the

TABLE 17.4 SUMMARY OF OUTPUT FROM FITTING ADDITIVE SEASONALITY TO THE AMTRAK RIDERSHIP DATA IN THE TRAINING PERIOD

> # include season as a predictor in tslm(). Here it creates 11 dummies, > # one for each month except for the first season, January. > train.lm.season <- tslm(train.ts ~ season) > summary(train.lm.season)

Call: lm(formula = formula, data = "train.ts", na.action = na.exclude)

Residuals: Min 1Q Median 3Q Max

-276.165 -52.934 5.868 54.544 215.081

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) 1573.97 30.58 51.475 < 2e-16 *** season2 -42.93 43.24 -0.993 0.3230 season3 260.77 43.24 6.030 2.19e-08 *** season4 245.09 44.31 5.531 2.14e-07 *** season5 278.22 44.31 6.279 6.81e-09 *** season6 233.46 44.31 5.269 6.82e-07 *** season7 345.33 44.31 7.793 3.79e-12 *** season8 396.66 44.31 8.952 9.19e-15 *** season9 75.76 44.31 1.710 0.0901 . season10 200.61 44.31 4.527 1.51e-05 *** season11 192.36 44.31 4.341 3.14e-05 *** season12 230.42 44.31 5.200 9.18e-07 *** --- Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 101.4 on 111 degrees of freedom Multiple R-squared: 0.6348, Adjusted R-squared: 0.5986 F-statistic: 17.54 on 11 and 111 DF, p-value: < 2.2e-16

A MODEL WITH TREND AND SEASONALITY 411

Amtrak ridership, the coefficient for season8 (396.66) indicates that the average number of passengers in August is higher by 396.66 thousand passengers than the average in January (the reference category). Using regression models, we can also capture multiplicative seasonality, where values in a certain season are on average, higher or lower by a percentage amount compared to another season. To fit multiplicative seasonality, we use the same model as above, except that we use log(Y ) as the outcome variable. In R, this is achieved by setting lambda=0 in the tslm() function.

17.3 A Model with Trend and Seasonality

Finally, we can create models that capture both trend and seasonality by including predictors of both types. For example, from our exploration of the Amtrak Ridership data, it appears that a quadratic trend and monthly seasonality are both warranted. We therefore fit a model to the training data with 13 predictors: 11 dummies for month, and t and t2 for trend. The fit and output from this final model are shown in Figure 17.6 and Table 17.5. If we are satisfied with

Time

R id

er sh

ip

14 00

20 00

26 00

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

Time

F or

ec as

t E rr

or s

− 40

0 0

40 0

1991 1993 1995 1997 1999 2001 2003 2005

Training Validation Future

FIGURE 17.6 REGRESSION MODEL WITH TREND AND SEASONALITY APPLIED TO AMTRAK RIDERSHIP (TOP) AND ITS FORECAST ERRORS (BOTTOM)

412 REGRESSION-BASED FORECASTING

TABLE 17.5 SUMMARY OF OUTPUT FROM FITTING TREND AND SEASONALITY TO AMTRAK RIDERSHIP IN THE TRAINING PERIOD

> train.lm.trend.season <- tslm(train.ts ~ trend + I(trend^2) + season) > summary(train.lm.trend.season)

Call: tslm(formula = train.ts ~ trend + I(trend^2) + season)

Residuals: Min 1Q Median 3Q Max

-213.77 -39.36 9.71 42.42 152.19

Coefficients: Estimate Std. Error t value Pr(>|t|)

(Intercept) 1696.9794 27.6752 61.32 < 0.0000000000000002 *** trend -7.1559 0.7293 -9.81 < 0.0000000000000002 *** I(trend^2) 0.0607 0.0057 10.66 < 0.0000000000000002 *** season2 -43.2458 30.2407 -1.43 0.1556 season3 260.0149 30.2423 8.60 0.00000000000006604 *** season4 260.6175 31.0210 8.40 0.00000000000018264 *** season5 293.7966 31.0202 9.47 0.00000000000000069 *** season6 248.9615 31.0199 8.03 0.00000000000126033 *** season7 360.6340 31.0202 11.63 < 0.0000000000000002 *** season8 411.6513 31.0209 13.27 < 0.0000000000000002 *** season9 90.3162 31.0223 2.91 0.0044 ** season10 214.6037 31.0241 6.92 0.00000000032920793 *** season11 205.6711 31.0265 6.63 0.00000000133918009 *** season12 242.9294 31.0295 7.83 0.00000000000344281 *** --- Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 70.9 on 109 degrees of freedom Multiple R-squared: 0.825, Adjusted R-squared: 0.804 F-statistic: 39.4 on 13 and 109 DF, p-value: <0.0000000000000002

this model after evaluating its predictive performance on the validation data and comparing it against alternatives, we would re-fit it to the entire un-partitioned series. This re-fitted model can then be used to generate k-step-ahead forecasts (denoted by Ft+k) by plugging in the appropriate month and index terms.

17.4 Autocorrelation and ARIMA Models

When we use linear regression for time series forecasting, we are able to account for patterns such as trend and seasonality. However, ordinary regression models do not account for dependence between values in different periods, which in cross-sectional data is assumed to be absent. Yet, in the time series context,

AUTOCORRELATION AND ARIMA MODELS 413

values in neighboring periods tend to be correlated. Such correlation, called autocorrelation, is informative and can help in improving forecasts. If we know that a high value tends to be followed by high values (positive autocorrelation), then we can use that to adjust forecasts. We will now discuss how to compute the autocorrelation of a series, and how best to utilize the information for improving forecasts.

Computing Autocorrelation

Correlation between values of a time series in neighboring periods is called auto- correlation, because it describes a relationship between the series and itself. To compute autocorrelation, we compute the correlation between the series and a lagged version of the series. A lagged series is a “copy” of the original series which is moved forward one or more time periods. A lagged series with lag-1 is the original series moved forward one time period; a lagged series with lag-2 is the original series moved forward two time periods, etc. Table 17.6 shows the first 24 months of the Amtrak ridership series, the lag-1 series and the lag-2 series.

Next, to compute the lag-1 autocorrelation, which measures the linear rela- tionship between values in consecutive time periods, we compute the correlation

TABLE 17.6 FIRST 24 MONTHS OF AMTRAK RIDERSHIP SERIES WITH LAG-1 AND LAG-2 SERIES

Month Ridership Lag-1 Series Lag-2 Series

Jan 91 1709 Feb 91 1621 1709 Mar 91 1973 1621 1709 Apr 91 1812 1973 1621 May 91 1975 1812 1973 Jun 91 1862 1975 1812 Jul 91 1940 1862 1975 Aug 91 2013 1940 1862 Sep 91 1596 2013 1940 Oct 91 1725 1596 2013 Nov 91 1676 1725 1596 Dec 91 1814 1676 1725 Jan 92 1615 1814 1676 Feb 92 1557 1615 1814 Mar 92 1891 1557 1615 Apr 92 1956 1891 1557 May 92 1885 1956 1891 Jun 92 1623 1885 1956 Jul 92 1903 1623 1885 Aug 92 1997 1903 1623 Sep 92 1704 1997 1903 Oct 92 1810 1704 1997 Nov 92 1862 1810 1704 Dec 92 1875 1862 1810

414 REGRESSION-BASED FORECASTING

between the original series and the lag-1 series (e.g., via the function cor()) to be 0.08. Note that although the original series in Table 17.6 has 24 time periods, the lag-1 autocorrelation will only be based on 23 pairs (because the lag-1 series does not have a value for January 1991). Similarly, the lag-2 autocorrelation, measuring the relationship between values that are two time periods apart, is the correlation between the original series and the lag-2 series (yielding −0.15).

We can use R’s Acf() function in the forecast package to directly compute and plot the autocorrelation of a series at different lags. For example, the output for the 24-month ridership is shown in Figure 17.7.

code for creating Figure 17.7

ridership.24.ts <- window(train.ts, start = c(1991, 1), end = c(1991, 24)) Acf(ridership.24.ts, lag.max = 12, main = "")

− 0.

6 −

0. 4

− 0.

2 0.

0 0.

2 0.

4 0.

6

Lag

A C

F

2 4 6 8 10 12

FIGURE 17.7 AUTOCORRELATION PLOT FOR LAGS 1–12 (FOR FIRST 24 MONTHS OF AMTRAK RIDERSHIP)

A few typical autocorrelation behaviors that are useful to explore are:

Strong autocorrelation (positive or negative) at a lag k larger than 1 and its multiples (2k, 3k, . . .) typically reflects a cyclical pattern. For exam- ple, strong positive lag-12 autocorrelation in monthly data will reflect an annual seasonality (where values during a given month each year are posi- tively correlated).

AUTOCORRELATION AND ARIMA MODELS 415

Positive lag-1 autocorrelation (called “stickiness”) describes a series where consecutive values move generally in the same direction. In the pres- ence of a strong linear trend, we would expect to see a strong and positive lag-1 autocorrelation.

Negative lag-1 autocorrelation reflects swings in the series, where high values are immediately followed by low values and vice versa.

Examining the autocorrelation of a series can therefore help to detect sea- sonality patterns. In Figure 17.7, for example, we see that the strongest autocor- relation is at lag 6 and is negative. This indicates a bi-annual pattern in ridership, with 6-month switches from high to low ridership. A look at the time plot confirms the high-summer low-winter pattern.

In addition to looking at the autocorrelation of the raw series, it is very useful to look at the autocorrelation of the residual series. For example, after fitting a regression model (or using any other forecasting method), we can examine the autocorrelation of the series of residuals. If we have adequately modeled the seasonal pattern, then the residual series should show no autocorrelation at the season’s lag. Figure 17.8 displays the autocorrelations for the residuals from the regression model with seasonality and quadratic trend shown in Figure 17.6. It is clear that the 6-month (and 12-month) cyclical behavior no longer dominates the series of residuals, indicating that the regression model captured them adequately. However, we can also see a strong positive autocorrelation from lag 1 on, indicating a positive relationship between neighboring residuals. This is valuable information, which can be used to improve forecasts.

− 0.

2 0.

0 0.

2 0.

4 0.

6

Lag

A C

F

2 4 6 8 10 12

FIGURE 17.8 AUTOCORRELATION PLOT OF FORECAST ERRORS SERIES FROM FIGURE 17.6

416 REGRESSION-BASED FORECASTING

Improving Forecasts by Integrating Autocorrelation Information

In general, there are two approaches to taking advantage of autocorrelation. One is by directly building the autocorrelation into the regression model, and the other is by constructing a second-level forecasting model on the residual series.

Among regression-type models that directly account for autocorrelation are autoregressive (AR) models, or the more general class of models called ARIMA (Autoregressive Integrated Moving Average) models. AR models are similar to linear regression models, except that the predictors are the past values of the series. For example, an autoregressive model of order 2, denoted AR(2), can be written as

Yt = β0 + β1Yt−1 + β2Yt−2 + ϵ (17.1)

Estimating such models is roughly equivalent to fitting a linear regression model with the series as the outcome variable, and the two lagged series (at lag 1 and 2 in this example) as the predictors. However, it is better to use designated ARIMA estimation methods (e.g., those available in R’s forecast package) over ordinary linear regression estimation, to produce more accurate results.4

Moving from AR to ARIMA models creates a larger set of more flexible fore- casting models, but also requires much more statistical expertise. Even with the simpler AR models, fitting them to raw time series that contain patterns such as trends and seasonality requires the user to perform several initial data transforma- tions and to choose the order of the model. These are not straightforward tasks. Because ARIMA modeling is less robust and requires more experience and sta- tistical expertise than other methods, the use of such models for forecasting raw series is generally less popular in practical forecasting. We therefore direct the interested reader to classic time series textbooks [e.g., s