A Variable Selection Method for Explainable AI using Hyponymy Relations of Knowledge Graph

Abstract

In recent years, there has been increasing interest in explainable AI. When designing a white-box machine learning model, it is important not only to increase the prediction accuracy of the model, but also to design a model with high interpretability. In this paper, we consider the case in which a knowledge graph is used as a resource for learning data, and propose a method for selecting explanatory variables with high interpretability by utilizing the hyponymy relations between entities. We further confirmed with a test using a dataset of professional baseball players that the method can indeed choose the desired explanatory variables.