On the role of relations and interactions in fuzzy systems for machine learning

Abstract

In this study, we address the challenge of explainability in machine learning within the context of multi-agent fuzzy systems. Specifically, we present two advances. On one hand, based on the Kruskal-Wallis test, we define the KWSHAP algorithm, which is applicable to any machine learning model. On the other hand, we introduce a method to define variables in a machine learning model when the input includes several fuzzy measures defined over a set of agents. This information is incorporated into the model through the calculation of the Shapley value and interaction indices. The KWSHAP algorithm and the subsequent graph representation allow us to interpret the role and importance of these interactions. Our findings enhance machine learning interpretability, offering a statistically robust and flexible framework for understanding relationships among features, thus contributing to more transparent and trustworthy decision-making processes