Shapley Lorenz Values for Artificial Intelligence Risk Management

Abstract

A trustworthy application of Artificial Intelligence requires to measure in advance its possible risks. When applied to regulated industries, such as banking, finance and insurance, Artificial Intelligence methods lack explainability and, therefore, authorities aimed at monitoring risks may not validate them. To solve this issue, explainable machine learning methods have been introduced to "interpret" black box models. Among them, Shapley values are becoming popular: they are model agnostic, and easy to interpret. However, they are not normalised and, therefore, cannot become a standard procedure for Artificial Intelligence risk management. This paper proposes an alternative explainable machine learning method, based on Lorenz Zonoids, that is statistically normalised, and can therefore be used as a standard for the application of Artificial Intelligence.

The empirical analysis of 15,000 small and medium companies asking for credit confirms the advantages of our proposed method.