PE-SHAP: Causally Interpretable Path-Wise Shapley Explanations

Maksym Buleshnyi, Joshua R. Loftus, Sakina Hansen

Track: Regular Track (Page limit: 6-8 pages)

Keywords: explainability, interpretability, shapley values, causality

TL;DR: Path-wise causal explanation framework for binary treatment settings with a particular focus on mediation.

Abstract: Explainability plays a critical role in ensuring that AI and machine learning models are transparent, trustworthy, and actionable - especially in high-stakes domains. However, many popular explanation techniques, such as Shapley values, focus on predictive rather than causal explanations. This limits their ability to inform decisions or policy. Recently, researchers have introduced variants of causally-aware Shapley values. In this paper, we extend a path-wise causal explanation framework for binary treatment settings, by introducing a new effect designed to better capture mediation. Additionally, we leverage doubly robust estimators to improve both reliability and efficiency. We validate our framework through simulations and real-world case studies, demonstrating its practical utility. We also show how individual-level explanations can be aggregated to estimate population-level effects, which allows broader causal analysis.

Submission Number: 10