Résumé
Interpretability of predictive machine learning models is crit-ical for numerous application contexts that require decisions to be un-derstood by end-users. It can be studied through the lens of local ex-plainability and attribution methods that focus on explaining a specificdecision made by a model for a given input, by evaluating the contri-bution of input features to the results, e.g. probability assigned to aclass. Many attribution methods rely on a game-theoretic formulationof the attribution problem based on an approximation of the popularShapley value, even if the underlying rationale motivating the use ofthis specific value is today questioned. In this paper we introduce theFESP - Fair-Efficient-Symmetric-Perturbation - attribution method asan alternative approach sharing relevant axiomatic properties with theShapley value, and the Equal Surplus value (ES) commonly applied incooperative games. Our results show that FESP and ES produce betterattribution maps compared to state-of-the-art approaches in image andtext classification settings.