Beyond Shapley Values: Cooperative Games for the Interpretation of Machine Learning Models

Shapley values, while popular for explaining AI models, may not be the best tool for all interpretability questions. The paper argues that the mathematical axioms justifying Shapley values don't necessarily align with what we actually need when trying to understand how machine learning models make decisions.

Broader alternatives exist in cooperative game theory that offer more flexibility. The Weber and Harsanyi sets represent families of allocation methods that can provide different but equally valid ways to attribute importance to features, allowing practitioners to choose approaches better suited to their specific interpretability goals.

There's an important distinction between how we measure a model's behavior and how we assign credit to individual features. The paper clarifies that "value functions" (measuring model outputs) and "aggregation rules" (deciding how to distribute credit) are separate choices, and conflating them can lead to misleading interpretations.

The field needs a more principled, systematic approach to designing explanation methods. Rather than adopting methods based on trends, the authors propose a three-step blueprint for constructing feature attributions that are theoretically grounded and robust, helping ensure explanations remain reliable as the field evolves.

One-size-fits-all explanation methods may not serve responsible AI well. By offering multiple theoretically justified approaches, organizations can select attribution methods aligned with their specific needs—whether that's regulatory compliance, user trust, or model debugging.