shap (shapley additive explanations)

SHAP (SHapley Additive exPlanations) attributes prediction to input features using game-theoretic Shapley values. **Core concept**: From cooperative game theory - fairly distribute "payout" (prediction) among "players" (features) based on their marginal contributions. **Properties**: Local accuracy (sum to prediction), missingness (zero contribution for absent features), consistency (larger contribution if feature has larger effect). **Computation**: Exact Shapley requires 2^n feature subsets - intractable. Approximations: KernelSHAP (sampling), TreeSHAP (efficient for tree models), DeepSHAP (deep learning). **For text**: Each token as feature, measure contribution to prediction. **Output interpretation**: Positive SHAP = pushes prediction higher, negative = pushes lower. Magnitude = importance. **Visualizations**: Force plots, summary plots, waterfall charts. **Advantages**: Theoretically grounded, consistent, model-agnostic. **Limitations**: Expensive for text (many tokens), baseline choice matters, correlations between features complicate interpretation. **Tools**: shap library (Python), extensive ecosystem. **Use cases**: Debug models, feature importance, model comparison, compliance explanations. Industry standard for explainability.