VARSHAP: A Variance-Based Solution to the Global Dependency Problem in Shapley Feature Attribution

ICLR 2026 Conference Submission19556 Authors

19 Sept 2025 (modified: 08 Oct 2025)ICLR 2026 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: explainability, Feature Attribution, Explainable AI (XAI), Shapley Values, Local Interpretability, Model-Agnostic Methods
Abstract: Feature attribution methods based on Shapley values, such as the popular SHAP framework, are built on strong axiomatic foundations but suffer from a critical, previously underappreciated flaw: global dependence. As recent impossibility theorems demonstrate, this vulnerability is not merely an estimation issue but a fundamental one. The feature attributions for a local instance can be arbitrarily manipulated by modifying the model's behavior in regions of the feature space far from that instance, rendering the resulting Shapley values semantically unstable and potentially misleading. This paper introduces VARSHAP, a novel feature attribution method that directly solves this problem. We argue that the source of the flaw is the characteristic function used in the Shapley game — the model's output itself. VARSHAP redefines this game by using the reduction of local prediction variance as the characteristic function. By doing so, our method is, by construction, independent of the model's global behavior and provides a truly local explanation. VARSHAP retains the desirable axiomatic properties of the Shapley framework while ensuring that the resulting attributions are robust and faithful to the model's local decision landscape. Experiments on synthetic and real-world datasets confirm our theoretical claims, showing that VARSHAP provides stable explanations under global data shifts where standard methods fail and demonstrates superior performance, particularly in robustness and complexity metrics.
Primary Area: interpretability and explainable AI
Submission Number: 19556
Loading