Learning interpretable multi-class models by means of hierarchical decomposition: Threshold Control for Nested Dichotomies

Abstract

The success of Artificial Intelligence at solving real-world problems poses the need for interpretable models, especially in human-centered applications. The multi-class scenario is often present in these environments; however, the majority of research on interpretability has focused on binary classification. In this work, a novel method based on hierarchical decompositions to obtain interpretable multi-class models is introduced. The proposal, named Threshold Control for Nested Dichotomies (TC-ND) method, creates a binary-based hierarchical class structure. Then, it discards meta-classes at each dichotomy of the structure according to a certain level of confidence, pursuing a modular and more comprehensible decomposition of the multi-class problem. The approach presents internal parameters that are optimized using the Differential Evolution algorithm. The goodness of our proposal is assessed using a twofold approach: performance is evaluated by comparing against other state-of-the-art multi-class methods; interpretability is supported by practical examples with counterfactual explanations and a discussion of the advantages that the TC-ND method presents regarding its transparency and auditability.