On the Connection between Game-Theoretic Feature Attributions and Counterfactual Explanations

Two popular explanation methods are mathematically connected: The paper proves that game-theoretic feature attributions (like SHAP, a widely-used tool for explaining AI decisions) and counterfactual explanations are theoretically equivalent under certain conditions, meaning they can produce the same insights about why an AI model made a particular decision.

Counterfactuals alone may not reliably show feature importance: The research reveals important limitations in using counterfactual explanations—which show "what would need to change for a different outcome"—to determine which features are most important to a model's decision. This matters because practitioners sometimes use counterfactuals for this purpose without realizing potential pitfalls.

Specific conditions determine when explanations are equivalent: The equivalency between these two explanation approaches only holds under particular mathematical conditions. Understanding these conditions helps practitioners know when they can trust one method versus the other, improving the reliability of AI explanations in practice.

The findings extend beyond one popular method: While the paper focuses on SHAP, the theoretical connection applies to broader game-theoretic solution concepts, suggesting the insights have wider applicability across different explanation techniques rather than being limited to a single tool.

Experimental validation confirms theoretical predictions: Testing on three datasets showed measurable differences in explanations at each stage of the connection, providing empirical support that the theoretical framework accurately describes how these explanation methods actually behave.