Interpreting Machine Learning Models

Description

Table of Contents About the Authors About the Technical Reviewers Acknowledgments Introduction Chapter 1: The Evolution of Machine Learning Defining Machine Learning The Evolution of Machine Learning Learning a Machine Learning Algorithm Piece It Together Focus on Specific Algorithm Descriptions Design an Algorithm Description Template Start Small and Build It Up Investigating Machine Learning Algorithm Behavior Step 1. Select an Algorithm Step 2. Identify a Question Step 3. Design the Experiment Step 4. Execute the Experiment and Report Results Step 5. Repeat What Does Machine Learning Model Accuracy Mean? Why Model Accuracy Is Not Enough Summary Chapter 2: Introduction to Model Interpretability Humans Are Explanation Hungry Explanations in Machine Learning What Are Black-Box Models? What Is Interpretability? The Motivation Behind Interpretability To Make Better Decisions To Eliminate Bias To Justify Processes To Reproduce Operations Displacement Strategy To Determine Practical Accuracy To Maintain Privacy To Understand Security Risks The Research Behind Interpretability Summary Chapter 3: Machine Learning Interpretability Taxonomy Scope-related Types of Post hoc Model Interpretability Global Model Interpretability on a Holistic Level Local Model Interpretability A Group of Predictions Model-related Types of Post hoc Model Interpretability Result-related Types of Post hoc Model Interpretability Categorizing Common Classes of Explainability Methods Summary Chapter 4: Common Properties of Explanations Generated by Interpretability Methods Explanation Defined Properties of Explanation Methods Template of Expression Transparency Mobility Algorithmic Feasibility Properties of Individual Explanations Correctness Loyalty Dependability Resoluteness Lucidness Reliability Significance Originality Representativeness Human-Friendly Explanations Contrastiveness Selectivity Social Focus on the Abnormal Truthful Consistent with Prior Beliefs General and Probable Summary Chapter 5: Human Factors in Model Interpretability Interpretability Roles Technical Expertise Builders Domain Knowledge Reviewers Stakeholders or End Users Interpretability Stages Ideation and Conceptualization Stage Building and Validation Stage Deployment, Maintenance, and Use Stage Interpretability Goals Interpretability for Model Validation and Improvement Interpretability for Decision-Making and Knowledge Discovery Interpretability to Gain Confidence and Obtain Trust Human-Friendly Themes Characterizing Interpretability Work Interpretability Is Cooperative Interpretability Is Process Interpretability Is a Mental Model Comparison Interpretability Is Context-Dependent Design Opportunities for Interpretability Challenges Identifying, Representing, and Integrating Human Expectations Communicating and Summarizing Model Behavior Scalable and Integrable Interpretability Tools Post-Deployment Support Summary Chapter 6: Explainability Facts: A Framework for Systematic Assessment of Explainable Approaches Explainability Facts List Dimensions Functional Requirements F1: Problem Supervision Level F2: Problem Type F3: Explanation Target F4: Explanation Breadth/Scope F5: Computational Complexity F6: Applicable Model Class F7: Relation to the Predictive System F8: Compatible Feature Types F9: Caveats and Assumptions Operational Requirements O1: Explanation Family O2: Explanatory Medium O3: System Interaction O4: Explanation Domain O5: Data and Model Transparency O6: Explanation Audience O7: Function of the Explanation O8: Causality vs. Actionability O9: Trust vs. Performance Usability Requirements U1: Soundness U2: Completeness U3: Contextfullness U4: Interactiveness U5: Actionability U6: Novelty U7: Complexity U8: Personalization Safety Requirements S1: Information Leakage S2: Explanation Misuse S3: Explanation Invariance Validation Requirements Summary Chapter 7: Interpretable ML and Explainable ML Differences Interpretable ML and Explainable ML Basics Analyzing the Decision Tree Digging Deeper Key Issues with Explainable ML Trade-offs Between Accuracy and Interpretability Beware of the Unfaithful Not Enough Detail Key Issues with Interpretable ML Profits vs. Losses Efforts to Construct Hidden Patterns Explanatory and Predictive Modeling Explaining or Predicting: The Key Differences Between Two Choices Validation, Model Evaluation, and Model Selection Validation Model Selection Model Use and Reporting Explanatory Models Summary Chapter 8: The Framework of Model Explanations Data Sets at a Glance Types of Frameworks for Tabular Data Feature Importance (FI) Predictive Power of Feature Subsets Additive Importance Measures Removal-based Explanations for Feature Importance Feature Removal Explaining Different Model Behaviors Summarizing Feature Influence Rule-based Explanations Prototypes Counterfactuals Explanations for Image Data Saliency Maps Concept Attribution Text Data Sentence Highlighting Attention-based Methods Summary Chapter 9: Feature Importance Methods: Details and Usage Examples Data Set Name Abstract Sources Data Set Information Attribute Information Random Forest Feature Importance Accuracy-based Importance Gini-based Importance Permutation Feature Importance Advantages Disadvantages Code SHAP Property 1 (Local Accuracy) Property 2 (Missingness) Property 3 (Consistency) SAGE How SHAP and SAGE Are Related LIME FACET Model Inspection Model Simulation Enhanced Machine Learning Workflow Code Synergy Redundancy Partial Dependence Plots (PDP) Code Individual Conditional Expectation DALEX Introduction to Instance-level Exploration Breakdown Plots for Additive Attributions Breakdown Plots for Interactions Ceteris Paribus Profiles Local Diagnostics Plots Implementation Example of DALEX on the Titanic Data Set Create a Pipeline Model Predict-level Explanations predict predict_parts predict_profile Model-level Explanations model_performance model_parts model_profile Summary Chapter 10: Detailing Rule-Based Methods MAGIE (Model-Agnostic Global Interpretable Explanations) MAGIE Algorithm Approach Preprocessing the Input Data Generating Instance Level Conditions Learning Rules from Conditions Postprocessing Rules Sorting Rules by Mutual Information GLocaLX Local to Global Explanation Problem Local to Global Hierarchy of Explanation Theories Finding Similar Theories Code Output Skope-Rules Methodology Implementation Anchors Finding Anchors Advantages Disadvantages Getting an Anchor Summary Chapter 11: Detailing Counterfactual Methods Counterfactual Explanations Use Case 1: Banking Software Use Case 2: Continuous Outcome Counterfactual Explanations at a Glance Generating Counterfactual Explanations Counterfactual Guided by Prototypes DiCE MOC (Multi-Objective Counterfactuals) Comparison Between the Algorithms DiCE Diversity and Feasibility Constraints Proximity Sparsity Optimization Advantages Disadvantages Summary Chapter 12: Detailing Image Interpretability Methods Image Interpretation Using LIME Step 1. Generate Random Perturbations for Input Image Step 2. Predict Class for Perturbations Step 3. Compute Weights (Importance) For the Perturbations Step 4. Fit an Explainable Linear Model Using the Perturbations, Predictions, and Weights Image Interpretation Using Pixel Attribution (Saliency Maps) Image Interpretation Using Class Activation Maps Step 1. Modify the Model Step 2. Retrain the Model with CAMLogger Callback Step 3. Use CAMLogger to See the Class Activation Map Step 4. Draw Conclusions from the CAM Image Interpretation Using Gradient-Weighted Class Activation Maps Summary Chapter 13: Explaining Text Classification Models Data Preprocessing, Feature Engineering, and Logistic Regression Model on the Data Interpreting Text Predictions with LIME Interpreting Text Predictions with SHAP Explaining Text Models with Sentence Highlighting Summary Chapter 14: The Role of Data in Interpretability Summary Chapter 15: The Eight Pitfalls of Explainability Methods Assuming One-Fits-All Interpretability Bad Model Generalization Unnecessary Use of Complex Models Ignoring Feature Dependence Interpretation with Extrapolation Confusing Linear Correlation with General Dependence Misunderstanding Conditional Interpretation Misleading Interpretations Due to Feature Interactions Misleading Feature Effects Due to Aggregation Failing to Separate Main from Interaction Effects Ignoring Model and Approximation Uncertainty Failure to Scale to High-Dimensional Settings Human-Intelligibility of High-Dimensional IML Output Computational Effort Unjustified Causal Interpretation Summary Conclusion Engage Interpretability with a Good Plan User Experience and Interpretability Go Hand in Hand Wherever Possible, Design the Model to Be Interpretable Choose Metrics to Reflect the End Goal and the End Task Understand the Trained Model Communicate Explanations to Model Users References Index