The Relationship between SHAP and Functional ANOVA

This repository demonstrates many of the ideas discussed in Statistical Aspects of SHAP: Functional ANOVA for Model Interpretation (Herren and Hahn 2022). Specifically, it:

1. Implements approximations of the functional ANOVA terms for a variety of reference distributions
2. Computes Shapley values as a weighted linear combination of functional ANOVA terms
3. Implements a breadth-first screening procedure (first introduced in Hooker (2004)) for identifying low-order functional ANOVA terms and speeding up SHAP convergence

References

Giles Hooker. Discovering Additive Structure in Black Box Functions. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 575–580, 2004.