Computing feature interactions 100× faster without losing accuracy

Understanding how machine learning models work requires knowing not just which features matter, but how they interact. Shapley and Banzhaf interactions measure this, but computing them exactly is exponentially slow. ProxySHAP uses lightweight proxy models (like decision trees) to approximate interactions cheaply, then corrects the bias mathematically—avoiding exponential computation while keeping errors small. Benchmarks show it outperforms existing methods by large margins on thousands-feature datasets, making explainability practical for real applications.