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

[4I1-GS-11-01] Constructing Decision Trees for Both Global and Local Explanations

〇Eisuke Kuroki¹, Satoshi Hara² (1. OSAKA UNIVERSITY, 2. The University of Electro-Communications)

Keywords:AI Explainability, Decision Tree, Local Explanation, Global Explanation

In global explanation methods for machine learning models, establishing criteria for selecting an appropriate explanation model from multiple candidates with similar predictive fidelity remains a challenge. In this study, we introduce explanation fidelity, which incorporates the consistency of local explanations, as a new evaluation criterion. We propose an explanation model that combines predictive trees and shallow allocation trees. Experiments using the Car Evaluation dataset demonstrate that the proposed method can improve explanation fidelity while maintaining a predictive fidelity of at least 0.95. Furthermore, we show that sufficient explanation performance can be achieved with approximately four predictive trees.