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

Abstract

Predictive models are increasingly used for managerial and operational decision-making. The use of complex machine learning algorithms, the growth in computing power, and the increase in data acquisitions have amplified the black-box effects in data science. Consequently, a growing body of literature is investigating methods for interpretability and explainability. We focus on methods based on Shapley values, which are gaining attention as measures of feature importance for explaining black-box predictions. Our analysis follows a hierarchy of value functions, and proves several theoretical properties that connect the indices at the alternative levels. We bridge the notions of totally monotone games and Shapley values, and introduce new interaction indices based on the Shapley-Owen values. The hierarchy evidences synergies that emerge when combining Shapley effects computed at different levels. We then propose a novel sensitivity analysis setting that combines the benefits of both local and global Shapley explanations, which we refer to as the “glocal” approach. We illustrate our integrated approach and discuss the managerial insights it provides in the context of a data-science problem related to health insurance policy-making.