Peter Bruce founded the Institute for Statistics Education at Statistics.com in 2002. The Institute specializes in introductory and graduate level online education in statistics, optimization, risk modeling, predictive modeling, data mining, and other subjects in quantitative analytics.

Grant Fleming is a Data Scientist at Elder Research Inc. (ERI). During his time at ERI, he has worked with clients in both government and the private sector on statistical testing, data asset creation, predictive analytics, and latent variable modeling. He has given multiple talks on machine learning interpretability and fairness within ERI as well as to outside groups. Internally to ERI, Grant is working on developing software packages for creating reproducible and interpretable black box models.

Introduction xix

Part I Motivation for Ethical Data Science and Background Knowledge 1

Chapter 1 Responsible Data Science 3

The Optum Disaster 4

Jekyll and Hyde 5

Eugenics 7

Galton, Pearson, and Fisher 7

Ties between Eugenics and Statistics 7

Ethical Problems in Data Science Today 9

Predictive Models 10

From Explaining to Predicting 10

Predictive Modeling 11

Setting the Stage for Ethical Issues to Arise 12

Classic Statistical Models 12

Black-Box Methods 14

Important Concepts in Predictive Modeling 19

Feature Selection 19

Model-Centric vs. Data-Centric Models 20

Holdout Sample and Cross-Validation 20

Overfitting 21

Unsupervised Learning 22

The Ethical Challenge of Black Boxes 23

Two Opposing Forces 24

Pressure for More Powerful AI 24

Public Resistance and Anxiety 24

Summary 25

Chapter 2 Background: Modeling and the Black-Box Algorithm 27

Assessing Model Performance 27

Predicting Class Membership 28

The Rare Class Problem 28

Lift and Gains 28

Area Under the Curve 29

AUC vs. Lift (Gains) 31

Predicting Numeric Values 32

Goodness-of-Fit 32

Holdout Sets and Cross-Validation 33

Optimization and Loss Functions 34

Intrinsically Interpretable Models vs. Black-Box Models 35

Ethical Challenges with Interpretable Models 38

Black-Box Models 39

Ensembles 39

Nearest Neighbors 41

Clustering 41

Association Rules 42

Collaborative Filters 42

Artificial Neural Nets and Deep Neural Nets 43

Problems with Black-Box Predictive Models 45

Problems with Unsupervised Algorithms 47

Summary 48

Chapter 3 The Ways AI Goes Wrong, and the Legal Implications 49

AI and Intentional Consequences by Design 50

Deepfakes 50

Supporting State Surveillance and Suppression 51

Behavioral Manipulation 52

Automated Testing to Fine-Tune Targeting 53

AI and Unintended Consequences 55

Healthcare 56

Finance 57

Law Enforcement 58

Technology 60

The Legal and Regulatory Landscape around AI 61

Ignorance Is No Defense: AI in the Context of Existing Law and Policy 63

A Finger in the Dam: Data Rights, Data Privacy, and Consumer Protection Regulations 64

Trends in Emerging Law and Policy Related to AI 66

Summary 69

Part II The Ethical Data Science Process 71

Chapter 4 The Responsible Data Science Framework 73

Why We Keep Building Harmful AI 74

Misguided Need for Cutting-Edge Models 74

Excessive Focus on Predictive Performance 74

Ease of Access and the Curse of Simplicity 76

The Common Cause 76

The Face Thieves 78

An Anatomy of Modeling Harms 79

The World: Context Matters for Modeling 80

The Data: Representation Is Everything 83

The Model: Garbage In, Danger Out 85

Model Interpretability: Human Understanding for Superhuman Models 86

Efforts Toward a More Responsible Data Science 89

Principles Are the Focus 90

Nonmaleficence 90

Fairness 90

Transparency 91

Accountability 91

Privacy 92

Bridging the Gap Between Principles and Practice with the Responsible Data Science (RDS) Framework 92

Justification 94

Compilation 94

Preparation 95

Modeling 96

Auditing 96

Summary 97

Chapter 5 Model Interpretability: The What and the Why 99

The Sexist Résumé Screener 99

The Necessity of Model Interpretability 101

Connections Between Predictive Performance and Interpretability 103

Uniting (High) Model Performance and Model Interpretability 105

Categories of Interpretability Methods 107

Global Methods 107

Local Methods 113

Real-World Successes of Interpretability Methods 113

Facilitating Debugging and Audit 114

Leveraging the Improved Performance of Black-Box Models 116

Acquiring New Knowledge 116

Addressing Critiques of Interpretability Methods 117

Explanations Generated by Interpretability Methods Are Not Robust 118

Explanations Generated by Interpretability Methods Are Low Fidelity 120

The Forking Paths of Model Interpretability 121

The Four-Measure Baseline 122

Building Our Own Credit Scoring Model 124

Using Train-Test Splits 125

Feature Selection and Feature Engineering 125

Baseline Models 127

The Importance of Making Your Code Work for Everyone 129

Execution Variability 129

Addressing Execution Variability with Functionalized Code 130

Stochastic Variability 130

Addressing Stochastic Variability via Resampling 130

Summary 133

Part III EDS in Practice 135

Chapter 6 Beginning a Responsible Data Science Project 137

How the Responsible Data Science Framework Addresses the Common Cause 138

Datasets Used 140

Regression Datasets—Communities and Crime 140

Classification Datasets—COMPAS 140

Common Elements Across Our Analyses 141

Project Structure and Documentation 141

Project Structure for the Responsible Data

Science Framework: Everything in Its Place 142

Documentation: The Responsible Thing to Do 145

Beginning a Responsible Data Science Project 151

Communities and Crime (Regression) 151

Justification 151

Compilation 154

Identifying Protected Classes 157

Preparation—Data Splitting and Feature Engineering 159

Datasheets 161

COMPAS (Classification) 164

Justification 164

Compilation 166

Identifying Protected Classes 168

Preparation 169

Summary 172

Chapter 7 Auditing a Responsible Data Science Project 173

Fairness and Data Science in Practice 175

The Many Different Conceptions of Fairness 175

Different Forms of Fairness Are Trade-Offs with Each Other 177

Quantifying Predictive Fairness Within a Data Science Project 179

Mitigating Bias to Improve Fairness 185

Preprocessing 185

In-processing 186

Postprocessing 186

Classification Example: COMPAS 187

Prework: Code Practices, Modeling, and Auditing 187

Justification, Compilation, and Preparation Review 189

Modeling 191

Auditing 200

Per-Group Metrics: Overall 200

Per-Group Metrics: Error 202

Fairness Metrics 204

Interpreting Our Models: Why Are They Unfair? 207

Analysis for Different Groups 209

Bias Mitigation 214

Preprocessing: Oversampling 214

Postprocessing: Optimizing Thresholds

Automatically 218

Postprocessing: Optimizing Thresholds Manually 219

Summary 223

Chapter 8 Auditing for Neural Networks 225

Why Neural Networks Merit Their Own Chapter 227

Neural Networks Vary Greatly in Structure 227

Neural Networks Treat Features Differently 229

Neural Networks Repeat Themselves 231

A More Impenetrable Black Box 232

Baseline Methods 233

Representation Methods 233

Distillation Methods 234

Intrinsic Methods 235

Beginning a Responsible Neural Network Project 236

Justification 236

Moving Forward 239

Compilation 239

Tracking Experiments 241

Preparation 244

Modeling 245

Auditing 247

Per-Group Metrics: Overall 247

Per-Group Metrics: Unusual Definitions of “False Positive” 248

Fairness Metrics 249

Interpreting Our Models: Why Are They Unfair? 252

Bias Mitigation 253

Wrap-Up 255

Auditing Neural Networks for Natural Language Processing 258

Identifying and Addressing Sources of Bias in NLP 258

The Real World 259

Data 260

Models 261

Model Interpretability 262

Summary 262

Chapter 9 Conclusion 265

How Can We Do Better? 267

The Responsible Data Science Framework 267

Doing Better As Managers 269

Doing Better As Practitioners 270

A Better Future If We Can Keep It 271

Index 273

The New copy of this book will include any supplemental materials advertised. Please check the title of the book to determine if it should include any access cards, study guides, lab manuals, CDs, etc.

The Used, Rental and eBook copies of this book are not guaranteed to include any supplemental materials. Typically, only the book itself is included. This is true even if the title states it includes any access cards, study guides, lab manuals, CDs, etc.