SDS Seminar Series – Drew Herren

The Fall 2023 SDS Seminar Series continues on November 10th from 2:00 p.m. to 3:00 p.m. with Drew Herren (Post Doc, University of Texas at Austin). This event is in-person.    

Title: Statistical Aspects of SHAP: Functional ANOVA for Model Interpretation

Abstract: SHAP is a popular method for measuring variable importance in machine learning models. In this work, we study the algorithm used to estimate SHAP scores and outline its connection to the functional ANOVA decomposition. We use this connection to show that challenges in SHAP approximations largely relate to the choice of a feature distribution and the number of functional ANOVA components identified. We argue that the parallels between machine learning explainability and sensitivity analysis are illuminating, but that the practical upshot of their relationship is that “silver bullet” methods for interpreting machine learning models are elusive and there is no alternative to in-depth “model interrogation.”