18 SHAP
shapshapley-valuesexplainabilityfeature-attributiontreeshap
Abstraction: SHAP method for explaining ML predictions via Shapley values
Key points:
- SHAP (SHapley Additive exPlanations) by Lundberg and Lee (2017) estimates Shapley values from coalitional game theory to fairly attribute each feature's contribution to a prediction
- Three estimation methods: KernelSHAP (model-agnostic, slow, O(TL2^M)), TreeSHAP (tree-specific, fast, O(TLD^2), the current default), and Permutation Method (efficient model-agnostic alternative)
- SHAP satisfies local accuracy, missingness, and consistency properties; connects LIME and Shapley values under a unified additive attribution framework
- KernelSHAP and permutation methods sample from the marginal distribution, ignoring feature dependence; conditional sampling changes the game and can assign non-zero values to unused features
- Global interpretation plots include feature importance (mean |SHAP|), summary plots, dependence plots, interaction values, and clustering by Shapley similarity
- Misleading SHAP explanations can be constructed intentionally (Slack et al. 2020), hiding model biases from explanation consumers
Connections: Shap · Shapley Values · Model Interpretability · Explainable AI
Source: https://christophm.github.io/interpretable-ml-book/shap.html