The many Shapley values for explainable artificial intelligence: A sensitivity analysis perspective

Emanuele Borgonovo, Elmar Plischke, Giovanni Rabitti

Abstract: Highlights•There is no one-size-fits-all approach to analyze a machine learning model.•We consider Shapley values for explaining black-box predictions at alternative levels.•We introduce new interaction indices based on Shapley-Owen values.•We propose a novel sensitivity analysis called the “glocal” approach.•Insights yielded by investigations at different scales are complementary rather than overlapping.