Source-linked AI summary
InterpretML: A Unified Framework for Machine Learning Interpretability
Harsha Nori, Samuel Jenkins, Paul Koch, Rich Caruana
TL;DR
InterpretML addresses the need for understandable machine-learning systems by unifying glassbox models and blackbox explanation methods in one extensible package. It combines a common API and visualization platform with the Explainable Boosting Machine, which is designed to retain intelligibility while achieving accuracy comparable to state-of-the-art methods.
Problem
Users need understandable machine-learning models and explanations for high-risk applications, debugging, compliance, and human-computer interaction.
Method
InterpretML unifies glassbox and blackbox interpretability algorithms through a compatible API, visualization platform, and the Explainable Boosting Machine.
Results
EBM is highly intelligible, often comparable in predictive power to Random Forest and XGBoost, and fast at prediction time despite added training cost.
Takeaways & Limitations
InterpretML supports comparing and extending interpretability methods, while EBM offers modular feature-level explanations for individual predictions and production deployment.
Abstract
from arXiv · showhide
InterpretML is an open-source Python package which exposes machine learning interpretability algorithms to practitioners and researchers. InterpretML exposes two types of interpretability - glassbox models, which are machine learning models designed for interpretability (ex: linear models, rule lists, generalized additive models), and blackbox explainability techniques for explaining existing systems (ex: Partial Dependence, LIME). The package enables practitioners to easily compare interpretability algorithms by exposing multiple methods under a unified API, and by having a built-in, extensible visualization platform. InterpretML also includes the first implementation of the Explainable Boosting Machine, a powerful, interpretable, glassbox model that can be as accurate as many blackbox models. The MIT licensed source code can be downloaded from github.com/microsoft/interpret.
1. Introduction
InterpretML addresses the growing need for understandable machine-learning models across high-risk and general applied settings. It unifies glassbox models and blackbox explanation methods with interactive comparison tools.
- Interpretability matters in healthcare, finance, judicial environments, model debugging, regulatory compliance, and human-computer interaction.
- InterpretML exposes state-of-the-art interpretability algorithms through a unified API.
- The framework covers glassbox models that are inherently intelligible and blackbox methods that explain opaque machine-learning pipelines.
- Interactive visualizations and a built-in dashboard support comparison among interpretability algorithms.
2. Package Design
InterpretML is organized around comparison, fidelity to reference methods, ecosystem compatibility, and modular use. Its unified API and visualization platform are presented through an architecture overview and example code.
- Ease of comparison: InterpretML makes algorithm comparison easier through a scikit-learn-style uniform API and a visualization platform centered on comparison.
- Stay true to the source: The package aims to expose reference algorithms and visualizations in their most accurate form.
- Play nice with others: InterpretML emphasizes compatibility with Jupyter Notebook, scikit-learn, and libraries including plotly, lime, shap, and SALib.
- Take what you want: Components can be used or extended independently, allowing computationally intensive explanations without the visualization layer and related dependencies.
- Figure 1 presents the code architecture, unified API, and relevant example code.
3. Explainable Boosting Machine
The Explainable Boosting Machine (EBM) is an interpretable generalized additive model designed to combine competitive predictive performance with feature-level explanations. It learns individual feature functions and optional pairwise interactions, while additive structure supports visualization and efficient prediction.
- EBM is a glassbox generalized additive model designed to achieve accuracy comparable to Random Forest and Boosted Trees while remaining intelligible and explainable.
- EBM learns each feature function with bagging and gradient boosting, restricting boosting to one feature at a time in round-robin order with a very low learning rate.This training design addresses feature collinearity and makes feature contributions individually interpretable.
- EBM can automatically detect and include pairwise interaction terms, further increasing accuracy while maintaining intelligibility.
- Because EBM is additive, each feature contributes modularly through a visualizable function f_j, making individual feature effects easier to understand.For an individual prediction, feature lookup contributions are added and passed through the link function to produce the final prediction.
- Figure 2 illustrates an increasing Age contribution to p(1) from ages 20 to 50 and a CapitalGain-dominated individual prediction.
- EBM often matches state-of-the-art predictive performance, incurs additional training cost, and executes predictions quickly using additions and feature-function lookups.Its light memory usage and fast prediction times make it attractive for production deployment.