Recently, the game theoretic solution concept called Shapley value gained attention in the machine learning community. It has been used to enhance explainability, specially to interpret results provided by machine learning models referred to as black box. For instance, in black box models, the impact of features towards performance measures are unclear. The Shapley value can be used as a model-agnostic approach to estimate such a contribution. In this study, we discuss both local and global approaches for machine learning interpretability based on Shapley values. We also present some recent applications and point out future perspectives. 

Retour page d'accueil du séminaire