Generalized Pref-SHAP to Explain Preference Functions

ICLR 2026 Conference Submission17822 Authors

Keywords: skew-symmetric, preference

TL;DR: A framework to explain generalized preference functions

Abstract: We address the problem of feature attribution for skew-symmetric preference functions in dueling data settings, using the cooperative game-theoretic concept of \textit{Shapley values}. Building on Pref-SHAP~\cite{hu2022explaining}, we propose \textit{Generalized Pref-SHAP}, a framework that extends its applicability to a broader class of preference functions. Our method leverages a simple neural network to model arbitrary feature mappings while exploiting the canonical block structure inherent to skew-symmetric functions, enabling more meaningful explanations. Additionally, we explore foundational questions about Pref-SHAP, including its relationship with the block decomposition structure of generalized preference functions. We perform experiments on a range of synthetic datasets to demonstrate the effectiveness and efficiency of our approach.

Primary Area: interpretability and explainable AI

Submission Number: 17822