Proxy-Based Approximation of Shapley and Banzhaf Interactions
Researchers have introduced ProxySHAP, a novel method for approximating Shapley and Banzhaf interactions in machine learning models. This approach combines the efficiency of tree-based proxy models with a residual correction technique to improve accuracy. ProxySHAP offers a polynomial-time generalization for calculating interaction indices in tree ensembles, overcoming previous limitations related to tree depth. Benchmarking shows ProxySHAP outperforms existing methods like ProxySPEX and KernelSHAP-IQ in approximation quality, even for large-scale applications with numerous features, and enhances downstream explainability tasks. AI
IMPACT Enhances explainability and approximation quality for complex ML models, potentially improving trust and debugging.