InterpretML

A glass-box model that is as accurate as state-of-the-art black-box models like Random Forest and Gradient Boosted Trees while remaining fully interpretable. EBM learns modular feature functions for each input, allowing exact, human-readable explanations of its predictions. Each feature's contribution is additive and can be visualized as a graph, making it possible to understand exactly how the model arrived at a specific decision. This is critical for fraud detection where analysts need to justify flagged transactions to regulators.

Provides a single, consistent interface for both training inherently interpretable models and explaining existing black-box models. This eliminates the need to learn multiple libraries for different interpretability tasks. The API supports classification and regression problems and integrates seamlessly with scikit-learn conventions. For fraud teams, this means you can train an EBM for transparent decisioning or use SHAP and LIME through the same interface to audit a deep learning model already in production.

Includes a built-in, interactive dashboard for exploring model behavior. Users can inspect global feature importance, drill down into local explanations for individual predictions, and analyze feature interaction strength. The dashboard visualizes how each feature contributes to a prediction across its entire range of values. For a fraud analyst reviewing a high-risk transaction, the dashboard can instantly surface the top reason codes—such as 'transaction amount 3x user average'—that drove the anomaly score.

Supports a suite of model-agnostic and model-specific explanation techniques for black-box models, including SHAP, LIME, Partial Dependence Plots, and Permutation Feature Importance. This allows teams to generate reason codes for any deployed model, even complex ensembles or neural networks. In financial fraud anomaly detection, this capability is essential for generating adverse action reason codes required by regulations when a transaction is blocked based on a model's output.

Beyond EBM, InterpretML supports training other inherently interpretable models such as Decision Trees, Logistic Regression, and RuleFit. RuleFit creates a sparse linear model combining simple, interpretable rules generated from decision trees with original features. This provides both predictive accuracy and complete transparency. For fraud detection, a RuleFit model can produce explicit, auditable rules like 'IF transaction_country != account_country AND amount > $10,000 THEN high_risk'.

Automatically detects and quantifies feature interaction strength using metrics like the H-statistic. This reveals non-additive behavior where the effect of one feature on the prediction depends on the value of another. In fraud detection, an interaction might exist between transaction time and merchant category, where late-night purchases at electronics stores carry a disproportionately higher risk. Understanding these interactions is vital for building robust, explainable fraud models.

InterpretML generates Partial Dependence Plots (PDP) and Individual Conditional Expectation (ICE) plots for global model understanding. PDPs show the average marginal effect of a feature, while ICE plots reveal heterogeneous effects across individual instances. Critical for fraud models:

• Visualizing how transaction amount affects risk score
• Detecting non-monotonic relationships that violate regulatory expectations
• Identifying where model behavior diverges across subpopulations

InterpretML supports both intrinsic interpretability (glass-box models like EBM) and post-hoc explainability (SHAP, LIME on black-boxes). Decision framework:

• Glass-box: Use when regulatory compliance requires exact, auditable logic (FCRA adverse action notices)
• Post-hoc: Use when maximum predictive performance is required and explanations are for internal debugging
• Hybrid: Train an EBM as a surrogate model to approximate a complex ensemble for global insight

InterpretML includes Morris sensitivity analysis for global feature importance screening. This method systematically varies one factor at a time across a grid of input values to compute elementary effects. Advantages over permutation importance:

• Computationally efficient for high-dimensional data
• Identifies features with non-linear effects and interactions
• Provides both mean and variance of feature influence
• Useful as a preprocessing step before building interpretable models