christophM/interpretable-ml-book

Features

• Model Explainability - Provides a comprehensive set of techniques for interpreting model decisions and quantifying feature contributions to predictions.
• Shapley Value Calculators - Provides tools and algorithms for computing Shapley values to interpret model predictions.
• Transparent Algorithm Design - Provides a comprehensive guide to building models with inherently interpretable structures.
• Black Box Model Approximation - Trains interpretable models to approximate and reveal the decision-making process of black-box systems.
• Explainable AI - Serves as a technical manual for interpreting model decisions and quantifying feature importance.
• Feature Contribution Analysis - Provides methods for estimating the influence of individual features on model predictions using game-theoretic marginal contributions.
• Feature Importance Attribution - Quantifies the relative contribution of individual input variables using Shapley values and permutation importance.
• Feature Influence Visualization - Plots how changes in a single feature affect a prediction while keeping other features constant.
• Feature Interaction Analyzers - Provides tools to analyze higher-order dependencies and interaction effects between input features.
• Attribution Gradients - Calculates feature importance by backpropagating model outputs to visualize influential input regions.
• Sparse Linear Model Fitting - Combines features and decision rules into a linear model using regularization to maintain sparsity.
• Local Surrogate Models - Trains simple interpretable models to approximate and explain the behavior of complex black-box predictions.
• Recursive Partitioning Trees - Implements decision trees that split data recursively to create transparent hierarchical prediction paths.
• ML Model Debugging Tools - Identifies shortcuts or errors in model behavior by detecting predictions that contradict domain knowledge.
• Model Behavioral Analysis - Describes average model behavior across a dataset using importance rankings and effect plots.
• Model Interpretability Tools - Provides a reference for implementing surrogate models and Shapley values to analyze black-box behavior.
• Interpretable Model Training - Details processes for fitting models that maintain human-understandable logic throughout the training phase.
• Perturbation-Based Sampling - Uses synthetic data neighbors created by altering feature values to probe model stability and contributions.
• Feature-Value Relationship Visualization - Produces plots mapping feature values to Shapley values to reveal input-output relationships.
• Counterfactual Explanation Generations - Identifies the minimal changes to input features required to change a model's prediction.
• Tabular Model Explanation - Includes analysis of how numerical and categorical features contribute to predictions, particularly for tabular datasets.
• Machine Learning Books - Provides a comprehensive textbook on techniques to make machine learning models transparent and understandable.
• Machine Learning Educational Resources - Offers theoretical and practical instructions for building inherently transparent models.
• Feature Influence Quantifications - Implements techniques to quantify how specific input variables influence model predictions.
• Additive-Component Decompositions - Implements techniques to decompose complex prediction functions into additive components to isolate feature influence.
• Concept - Maps inputs to a set of concepts in a bottleneck layer to enable counterfactual explanations.
• Concept-Bottleneck Mappings - Details how to map high-dimensional inputs to human-understandable concepts to improve model transparency.
• Evaluation Datasets - Supplies processed regression and classification datasets used for testing the transparency of models.
• Representative Subset Selection Algorithms - Provides algorithms for selecting diverse and representative subsets from datasets to preserve distribution coverage.
• Decision List Construction - Details the construction of decision lists that iteratively explain datasets through sequential rules.
• Decision Trees - Implements decision tree models that map observations to target values via recursive splitting.
• Neuron Activation Visualization - Identifies inputs that maximize the activation of specific neurons to reveal learned visual patterns.
• Dependent Feature Decomposition - Uses functional ANOVA to decompose models with correlated features using hierarchical orthogonality.
• Domain Insight Extraction - Provides methodologies to uncover scientific patterns or business drivers by analyzing relationships between input features and targets.
• Conditional Feature Importance - Evaluates a feature's unique contribution by sampling from conditional distributions to account for dependencies.
• Feature Interaction Models - Implements models that capture combined effects between features through interaction terms and varying slopes.
• Interaction Effect Calculation - Computes numerical values for feature interactions to reveal hidden dependencies in model predictions.
• Causal Effect Decompositions - Produces a functional decomposition to isolate feature effects without entangling correlated features.
• Second-Order Interaction Analyses - Identifies second-order effects by isolating interaction between two features from their main effects.
• Frequent Pattern Extraction - Implements techniques to identify common combinations of feature values used to establish decision rules.
• Functional ANOVA Execution - Decomposes a function into orthogonal components to isolate interaction effects from main effects.
• Interpretability Assessments - Assesses explanation quality using tests with end users, laypersons, or proxy tasks.
• Logistic Regression Weight Interpretations - Translates model weights into odds ratios to determine how features affect binary outcomes.
• L1 Regularization - Uses L1 regularization to constrain the number of active features, ensuring human-readable model complexity.
• Interpretability-Based Debugging - Identifies shortcuts or errors in model behavior by comparing predictions against domain knowledge and data distributions.
• Marginal Effect Visualizations - Plots the average relationship between features and outcomes in the project by marginalizing over other distributions.
• Glassbox Model Construction - Provides guidance on constructing models that are inherently transparent and interpretable by design.
• Training Data Influence Analysis - Identifies which training examples most influenced specific predictions to debug model behavior.
• Concept Attribution Methods - Quantifies how much a human-defined concept influences a model's prediction for a particular class.
• Functional Decompositions - Breaks complex models into a sum of additive components and interaction terms for transparency.
• Model Weight Visualizations - Generates plots of coefficients and confidence intervals to identify significant prediction drivers.
• Neural Network Interpretability - Links activated neural network channels with human-understandable concepts to measure detection quality.
• Nonlinear Relationship Modeling - Fits non-linear patterns using feature transformations or smooth curves to represent complex real-world trends.
• Explanation Property Analysis - Measures the effectiveness of explanations based on their expressive power and algorithmic complexity.
• Prediction Justification Frameworks - Provides frameworks for explaining the reasoning behind specific model outputs to ensure transparency and recourse.
• Prediction-Specific Rule Generators - Provides techniques for generating rules that isolate the specific feature values driving a particular classification.
• Architecture Interpretation - Analyzes internal architecture, weights, and gradients to uncover learned concepts within models.
• Saliency Map Localizations - Generates coarse heatmaps by backpropagating gradients in the project to identify regions driving a classification.
• Pixel Saliency Maps - Calculates gradients of a class score to visualize image regions influencing a prediction.
• Underrepresented Data Point Identification - Finds data instances poorly represented by prototypes to reveal gaps in data coverage.
• Bayesian Rule List Learning - Implements Bayesian rule list learning to create highly interpretable decision logic.
• Global Impact Rankings - Ranks features by averaging absolute Shapley values to determine global model impact.
• Tree Model Interpretability - Provides exact attribution methods for decision trees and ensemble models like Random Forests or XGBoost.
• Visual Concept Discovery - Clusters image segments to automatically identify influential visual concepts within a dataset.
• Tree Ensemble Rule Extraction - Implements the extraction of binary interaction rules from complex tree ensembles.
• Leave-One-Feature-Out Importance - Measures feature contributions by retraining the model without specific features and comparing performance.
• Explanation Fidelity Testing - Validates the faithfulness of model explanations by checking accuracy, fidelity, and consistency.
• Individual Conditional Expectation Curves - Provides visualizations showing how predictions change for individual data points as specific features vary.
• Partial Dependence Importance Metrics - Quantifies feature importance by measuring the variance of its partial dependence curve.

Star history