[4I1-GS-11] AI and Society:

[4I1-GS-11-02] A study on fuzzy classifier design considering interpretability and fairness by a quality diversity algorithm

〇Takeru Konishi¹, Naoki Masuyama¹, Yusuke Nojima¹ (1. Osaka Metropolitan University)

Keywords:Fairness, Interpretability, Evolutionary Computation, Quality Diversity Algorithm, Fuzzy System

There is a growing interest in various aspects of artificial intelligence, beyond accuracy, to address the ethical and social risks. Among them, transparency and fairness are important, and considering fairness in highly transparent models has attracted much attention. Especially, considering fairness in intrinsically interpretable models is expected to be used in cases where transparency and fairness are required as it enables us to consider fairness together with an understanding of the internal mechanism of the model. Fuzzy classifiers are representative intrinsically interpretable models that can make decisions taking into account real-world uncertainties. In this study, we study the fuzzy classifier design by MAP-Elites, a representative quality diversity algorithm, considering interpretability and fairness. We show that by using MAP-Elites, which can improve the performance together with the diversity of selected features, it is possible to obtain a highly accurate set of fuzzy classifiers with high diversity in terms of interpretability and fairness. Furthermore, we discuss the relationship between accuracy, interpretability, and fairness using the feature-performance map.