On measuring features importance in Machine Learning models in a two-dimensional representation scenario

Abstract

Abstract: There is a wide range of papers in the literature about the explanation of machine learning models in which Shapley value is considered to measure the importance of the features in these models. We can distinguish between these which set their basis on the cooperative game theory principles, and these focused on fuzzy measures. It is important to mention that all of these approaches only provide a crisp value (or a fix point) to measure the importance of a feature in a specific model. The reason is that an aggregation process of the different marginal contributions produces a single output for each variable. Nevertheless, and because of the relations between features, we cannot distinguish the case in which we do not know all the features. To this aim, we propose a disaggregated model which allows the analysis of the importance of the features, regarding the available information. This new proposal can be viewed as a generalization of all previous measures found in literature providing a two dimensional graph which, in a very intuitive and visual way, provides this rich disaggregated information. This information may be aggregated with several aggregation functions with which obtain new measures to establish the importance of the features. Specifically, the aggregation by the sum results in the Shapley value. We also explain the characteristics of those graphics in different scenarios of the relations among features, to raise this useful information at a glance.