Coalitional Explanations

In this work on Explainable AI, I created a new explanation method for the popular game-theory based model agnostic explainability library SHAP. My work cleared up some theoretical confusion interpreting hierarchical Shapley values and created an efficient principled way to explain any machine learning model using feature coalition structures. These coalitional explanations deriving from Winter values can provide a computationally efficient approximation of Shapley values consistent with the correlation or model structure of features.