On the application of the choquet integral in the context of explainable artificial intelligence

Abstract

A central issue in the study of machine learning method regards the search for interpretability in models obtained after a training phase. Indeed, in additional to providing to the decision maker elements about how the model is using the information available to accomplish a given task, interpretability is a crucial requirement for more transparent algorithms. Motivated by the current efforts towards interpretable models, which falls within the more general scope of Explainable Artificial Intelligence (XAI), this project aims at investigating the application of a model known as Choquet integral in the scope of XAI. We shall study, through numerical experiments, if the Choquet integral is able to achieve a good performance both in terms of accuracy and interpretability. Moreover, we shall compare the Choquet integral with more classic approaches often adopted to tackle problems such as regression and classification.