Stratum by Stratum: Building Stable SHAP Explanations through Layered Approximations
Résumé
SHAP is a popular post-hoc explainability method that assigns feature attribution scores based on the Shapley value from cooperative game theory. Since computing the exact SHAP values is intractable for large feature sets, several sampling-based approximation methods, such as KernelSHAP, have been proposed in the literature. These methods overcome intractability but suffer from stability issues. To address this instability, we propose the StratoSHAP family of stable SHAP approximations based on feature coalitions organized into strata. We provide formal analytical formulations for these approximations and demonstrate that they respect important properties of attribution values, such as stability, linearity, efficiency, symmetry, and fair treatment. A series of comparative experiments reveal that our stratum-based approach offers an interesting trade-off between computational complexity and approximation quality while remaining fully stable.
Origine | Fichiers produits par l'(les) auteur(s) |
---|