Application of Fuzzy Classifiers to Hybrid Interpretable Models

Abstract

Interpretability is an important factor in real-world AI applications. There are mainly two research directions. One is to give explanations to black-box models in a post-hoc manner, and the other is to directly use interpretable white-box models. While each approach has been actively studied, combining the two approaches has not been discussed well. This paper proposes a hybrid interpretable model by combining a fuzzy classifier (i.e., white-box model) and a random forest classifier (i.e., black-box model). The fuzzy classifier gives linguistic rules for as many patterns as possible, while the random forest classifier gives accurate predictions for difficult patterns. Computational experiments using real-world datasets suggested that the proposed method is promising for both classification performance and interpretability.