Shapley values for feature selection: The good, the bad, and the axioms

Shapley values have theoretical limitations for feature selection despite their popularity in explainable AI: While Shapley values are mathematically elegant and satisfy four axioms that make them theoretically sound for attribution in games, the paper demonstrates through counterexamples that these same axioms can actually work against the goals of feature selection.

The axioms that make Shapley values "fair" may select the wrong features: The paper shows that features satisfying the Shapley value axioms (such as symmetry and null player properties) may not be the most important ones for prediction or understanding a model, suggesting a fundamental mismatch between theoretical fairness and practical feature importance.

Popular implementations like SHAP and SAGE inherit these limitations: The research investigates concrete implementations of Shapley-based methods (SHAP and SAGE) in simulation settings and finds that the theoretical issues translate to real-world problems, meaning users of these widely-adopted tools should be aware of their constraints.

Feature selection requires different criteria than game-theoretic fairness: The paper suggests that selecting important features for understanding AI models may require abandoning some of the axioms that make Shapley values theoretically appealing, indicating that practitioners may need alternative approaches or modified methods for this specific use case.