Presentation Information
[1K4-GS-3b-05]Interpretable Prediction Model Based on Genetic Network Programming with Discovered Rules as Nodes
〇Shuhei Tanaami1, Kaoru Shimada1 (1. Graduate School of Informatics, Gunma University)
Keywords:
Explainable AI,Rule-based learning,Evolutionary computation,Data mining
Although high-performance machine learning methods, such as deep learning, achieve high prediction accuracy, their complex model structures make it difficult for humans to understand the basis of the prediction results. In contrast, rule-based methods can express decision criteria as explicit rules and have high interpretability. However, conventional methods that simultaneously use multiple rules make it difficult to understand which rules are applied and how they are applied. This study aims to build an interpretable prediction model that can trace the decision-making process at the rule level while maintaining the prediction accuracy. To this end, we propose a prediction network based on genetic network programming (GNP). Numerous rules discovered by previous research methods are placed as GNP judgement nodes, and stepwise decision-making is performed through network transitions, simultaneously optimizing the rule selection, combination, connection structure, and application order. Furthermore, high interpretability was ensured by limiting the number of rules applied to each prediction to a maximum of three. Experiments using a geomusic dataset, which is difficult to predict with conventional machine learning models, demonstrated that the proposed method can clearly demonstrate the rule application process while maintaining prediction accuracy.