interpret
Fit interpretable models. Explain blackbox machine learning.
InterpretML is a Microsoft open-source Python package for building explainable machine learning models and understanding why any existing model made a specific prediction, including a high-accuracy transparent model called EBM.
InterpretML is an open-source Python package from Microsoft that addresses a common frustration with machine learning: models that make good predictions but give no explanation for how they reached a decision. The package offers two main paths for dealing with this. First, you can train models that are inherently transparent, meaning their logic can be read and checked directly. Second, you can take an existing model of any kind and use explanation techniques to understand why it produced a particular result.
The centerpiece of the package is a model type called the Explainable Boosting Machine, developed at Microsoft Research. It is designed to match the accuracy of popular but opaque approaches like random forests and gradient boosted trees while also producing explanations that a human can inspect and, if needed, adjust. Benchmark results in the README show it performing at or near the top across finance, medical, and business datasets. For situations where the training data must stay private, a differentially private variant is also available that provides formal guarantees about how much information about individual records could be inferred from the model.
Beyond its own model type, InterpretML integrates existing explanation methods like SHAP and LIME, which can be applied to any model, treating it as a black box and approximating why it made each prediction. Partial dependence plots and sensitivity analysis are also included.
The package provides a dashboard interface for comparing explanations side by side across multiple models, making it easier to spot where models agree or disagree in their reasoning. It works with standard Python data formats and integrates with the scikit-learn ecosystem, so it slots into existing data science workflows without much friction.
Installation is a single pip or conda command and requires Python 3.10 or higher. The package runs on Linux, macOS, and Windows. The README lists common reasons someone might need model interpretability: debugging errors, checking for bias or discrimination, satisfying regulations, or building trust when the stakes are high. The full README is longer than what was shown.
Where it fits
- Train an Explainable Boosting Machine to get predictions as accurate as a random forest with human-readable explanations you can inspect and adjust.
- Audit an existing black-box model for bias or discrimination using SHAP or LIME explanation methods.
- Compare multiple models side by side in the InterpretML dashboard to see where they agree or disagree in their reasoning.
- Train a differentially private EBM when your training data is sensitive and formal privacy guarantees are required.