InterpretML — Interpretable Machine Learning by Microsoft

Introduction

InterpretML is an open-source library from Microsoft Research that unifies interpretable (glass-box) models and explainability techniques (black-box) under a single API. Its flagship model, the Explainable Boosting Machine (EBM), achieves accuracy comparable to gradient boosting while remaining fully interpretable.

What InterpretML Does

• Trains glass-box models like EBM, linear models, and decision trees
• Explains any black-box model with SHAP, LIME, and partial dependence
• Provides interactive visualizations of feature importances and effects
• Supports both classification and regression tasks
• Enables comparison of multiple explanations side by side

Architecture Overview

InterpretML defines an Explainer interface with explain_global and explain_local methods. Glass-box models implement both training and explanation natively. Black-box explainers wrap external models and produce explanations via perturbation or gradient methods. All explanations are rendered as interactive Plotly dashboards.

Self-Hosting & Configuration

• Install via pip; the package includes all glass-box models
• Use ExplainableBoostingClassifier or Regressor as drop-in sklearn estimators
• Set max_bins and interactions to control EBM complexity
• For black-box explanations, wrap any predict function with ShapKernel or LimeTabular
• Launch the interactive dashboard with show() in a Jupyter environment

Key Features

• EBM matches gradient boosting accuracy with full interpretability
• Unified API across glass-box and black-box explanation methods
• Interactive Plotly-based dashboards for exploring feature effects
• Pairwise interaction detection built into the EBM training loop
• Differential privacy support for training on sensitive data

Comparison with Similar Tools

• SHAP — focuses on Shapley value explanations for any model; InterpretML includes SHAP plus its own glass-box models
• LIME — local explanation technique; InterpretML integrates LIME alongside other methods
• Alibi — strong on counterfactual and anchor explanations; InterpretML focuses on feature-level interpretation
• Captum — PyTorch-specific attribution methods; InterpretML is framework-agnostic

FAQ

Q: What is an Explainable Boosting Machine? A: EBM is a generalized additive model with pairwise interactions trained via cyclic gradient boosting. Each feature's effect is learned independently, making the model fully inspectable.

Q: Does EBM work with large datasets? A: Yes. EBM scales well and supports parallel training via the n_jobs parameter.

Q: Can I use InterpretML with deep learning models? A: Yes. The black-box explainers work with any model that has a predict or predict_proba method.

Q: Is InterpretML production-ready? A: Yes. EBM models can be serialized with joblib or pickle and served like any scikit-learn model.

Sources

• https://github.com/interpretml/interpret
• https://interpret.ml/docs