Cooperative Games for the Interpretation of Machine Learning Models

Abstract: Cooperative game theory has become a cornerstone of post-hoc interpretability in machine learning, largely through the use of Shapley values. Yet, despite their widespread adoption, Shapley-based methods often rest on axiomatic justifications whose relevance to feature attribution remains debatable. During this presentation, we will revisit cooperative game theory from an interpretability perspective and argue for a more principled use of its tools. Through two broad families of allocations, we will provide an intuitive interpretation of the Shapley values, offer a blueprint for defining more intricate interpretation tools, and derive statistically sound estimates to solve the exponential computational burden. Finally, we will discuss the theoretical challenges surrounding the choice of value function.