[Paper Review] InterpretML: A Unified Framework for Machine Learning Interpretability

Harsha Nori, Samuel Jenkins+2 more|arXiv (Cornell University)|Sep 19, 2019

Explainable Artificial Intelligence (XAI)16 references316 citations

TL;DR

InterpretML provides a unified Python framework for machine learning interpretability, combining glassbox models with blackbox explainability tools and introducing the Explainable Boosting Machine (EBM).

ABSTRACT

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.

Motivation & Objective

• Provide a unified API to compare interpretability algorithms across glassbox and blackbox approaches.
• Expose interpretable models (glassbox) and model-agnostic explanations (blackbox) via a consistent interface.
• Enable visualization and dashboard-based comparison to aid model interpretability decisions.
• Introduce and evaluate the Explainable Boosting Machine (EBM) as an accurate and intelligible model.

Proposed method

• Adopts a scikit-learn–style API to ease comparison of interpretability algorithms.
• Presents two interpretability forms: glassbox models (inherently interpretable) and blackbox explanations (for any pipeline).
• Introduces the Explainable Boosting Machine (EBM), a generalized additive model that learns feature functions and optional pairwise interactions.
• Uses round-robin boosting over features with small learning rate to mitigate multicollinearity and ensure additive interpretability.
• EBM is implemented in C++/Python with parallelization via joblib for scalable training and fast predictions.
• Provides visualization of feature contributions f_j(x_j) to interpret individual predictions.

Experimental results

Research questions

• RQ1 How can interpretability algorithms be exposed under a unified API to facilitate comparison?
• RQ2 Can a highly interpretable glassbox model (EBM) achieve competitive predictive performance with state-of-the-art blackbox methods?
• RQ3 What are the computational characteristics (training/prediction) of EBMs compared to other models across diverse datasets?
• RQ4 How effective are visualization and interaction tools for understanding feature contributions and interactions?

Key findings

• EBM often achieves predictive performance comparable to state-of-the-art models such as Random Forest and XGBoost across multiple datasets.
• EBM provides fast predictions with low memory usage due to its additive structure and simple term lookups.
• EBM supports automatic inclusion of pairwise interactions to boost accuracy while preserving interpretability.
• The framework enables easy comparison of multiple interpretability algorithms and includes interactive visualizations and a dashboard.
• EBM default parameters emphasize speed, with recommended reference parameters for best accuracy and interpretability.