shap values

SHAP values (SHapley Additive exPlanations) is a feature attribution method that applies Shapley values from cooperative game theory to machine learning models. For a given prediction, each feature receives a SHAP value representing its additive contribution relative to a baseline (the expected model output over a background dataset). The Shapley value is computed by considering all possible coalitions of features, weighted by the number of possible orderings — ensuring the attribution satisfies three axioms: local accuracy (the prediction equals the sum of SHAP values plus the baseline), missingness (a missing feature gets zero attribution), and consistency (if a model relies more on a feature, its SHAP value weakly increases). In practice, exact computation is intractable for most models with many features, so approximations like TreeSHAP (for tree ensembles) or DeepSHAP (for neural networks) are used. The method is model-agnostic in principle but has efficient implementations for specific model families. [formal: shap | substrate: behavior | horizon: hours | explicit: yes | epoch: 0.01]