shap

Unlike simpler feature importance methods that describe global model behavior, SHAP provides local explanations for individual predictions: exactly how much did each feature push this prediction up or down from the base rate, with mathematically provable properties including consistency, local accuracy, and missingness.

The library provides optimized SHAP implementations for different model types: TreeExplainer for tree-based models (XGBoost, LightGBM, random forests) that computes exact Shapley values in polynomial time using the tree structure, DeepExplainer for neural networks using a gradient-based approximation, LinearExplainer for linear models with exact closed-form solutions, and KernelExplainer as a model-agnostic method for any black-box model.

Visualization toolswaterfall plots, beeswarm plots, force plots, dependency plotscommunicate SHAP values in forms accessible to stakeholders without technical ML background.

Data scientists explaining model predictions to business stakeholders, compliance teams satisfying right-to-explanation requirements under GDPR and similar regulations, ML engineers debugging models by understanding which features drive unexpected predictions, and practitioners studying feature importance for feature engineering decisions use SHAP as the standard tool for machine learning interpretability.

Its theoretical grounding distinguishes it from ad-hoc explanation methods, and its consistent results across model types make it a reliable framework for fairness auditing and model validation workflows.