SHAP

SHAP (SHapley Additive exPlanations) is the gold-standard open-source library for machine learning model explainability, providing a unified approach to interpreting model predictions based on the mathematically rigorous framework of Shapley values from cooperative game theory. Developed by Scott Lundberg and collaborators at the University of Washington, SHAP has become the most widely adopted and cited explainability tool in the machine learning community, with applications spanning virtually every industry and model type. At its core, SHAP assigns each feature an importance value for a particular prediction, indicating how much that feature contributed to pushing the prediction away from the base value (the average model prediction). Unlike ad hoc feature importance methods, SHAP values satisfy three desirable theoretical properties: local accuracy (the explanation matches the model prediction), missingness (features with no impact receive zero attribution), and consistency (a feature's attribution never decreases when the model relies more on it). These properties make SHAP the most theoretically grounded approach to feature attribution in machine learning. The library provides optimized implementations for several model types, including TreeSHAP for tree-based models (XGBoost, LightGBM, Random Forests), DeepSHAP for deep learning models, and KernelSHAP as a model-agnostic fallback. TreeSHAP is particularly noteworthy for its computational efficiency, computing exact Shapley values in polynomial time for tree ensembles, whereas the general Shapley computation is exponential. SHAP provides rich visualization capabilities including force plots, summary plots, dependence plots, and interaction plots that can communicate model behavior to both technical and non-technical audiences. These visualizations have become standard in model documentation, regulatory submissions, and stakeholder presentations across regulated industries. The library integrates seamlessly with the Python data science stack including scikit-learn, XGBoost, LightGBM, TensorFlow, and PyTorch. It is freely available under the MIT license and has garnered over 20,000 GitHub stars, making it one of the most popular machine learning libraries. However, SHAP is a developer library requiring Python programming expertise and can be computationally expensive on large datasets, particularly when using the model-agnostic KernelSHAP approach.