Resolving the Duplicate-Feature Paradox with ReSHAP: A Redundancy-Weighted Generalization of Shapley Attribution

ICLR 2026 Conference Submission21340 Authors

Keywords: Explainable AI (XAI), Shapley values, Feature attribution, Redundancy-aware explanations, Mutual information, Cooperative game theory, Model interpretability, Feature importance, Value function decomposition, Duplicate-feature paradox

TL;DR: ReSHAP is a redundancy-aware generalization of Shapley attribution that resolves the duplicate-feature paradox by down-weighting redundant features, ensuring fairer and computationally efficient feature attributions in machine learning explanations

Abstract:

Feature attribution methods based on Shapley values are widely used to explain machine learning model predictions. However, these methods suffer from a critical flaw often observed when features are duplicated, which we term the duplicate-feature paradox: when a feature is duplicated (or strongly correlated with another), its total contribution to the model prediction is unfairly inflated, diminishing the attribution of other important features. This paradox arises because traditional Shapley-based methods allocate joint contributions equally across all participating features, regardless of redundancy or informational overlap.

In this work, we propose ReSHAP, a redundancy-aware generalization of Shapley attribution that systematically resolves the duplicate-feature paradox. ReSHAP adjusts the allocation of credit within feature coalitions by down-weighting features that contribute redundant information. We begin by proving that no attribution method can simultaneously satisfy equal division and duplication-invariance, even in instances without redundant features. This reveals a fundamental trade-off in designing fair attribution methods. Building on this insight, ReSHAP redefines how Shapley values are computed by redistributing interaction terms across feature subsets using a recursive weighting scheme. This approach preserves core theoretical properties while providing a practical, computationally efficient correction for redundancy bias, without altering the standard value function or requiring distributional assumptions. We support our theoretical findings with illustrative examples and experiments, highlighting the practical effectiveness of ReSHAP.

Primary Area: interpretability and explainable AI

Submission Number: 21340