Evolutionary Multi-Objective Multi-Tasking for Fuzzy Genetics-Based Machine Learning in Multi-Label Classification

Yuichi Omozaki, Naoki Masuyama, Yusuke Nojima, Hisao Ishibuchi

Abstract: Explainable artificial intelligence (XAI) is an important research topic in the field of machine learning. A fuzzy rule-based classifier is a promising XAI technique thanks to its high interpretability. We can linguistically explain its classification result because a set of linguistically explainable fuzzy if-then rules are used for classification. In real-world data mining applications, multiple class labels are assigned to a single instance. Such a dataset is called a multi-label dataset (MLD). For MLDs, multiobjective fuzzy genetics-based machine learning for multi-label classification (MoFGBMLML) has been proposed. MoFGBMLML aims to search for explainable fuzzy classifiers by explicitly considering the accuracy-complexity tradeoff that exists in explainable classifier design. In the field of multi-label classification, different accuracy metrics have been proposed to evaluate classifier performance. As a result, different multiobjective optimization problems (MOPs) can be defined using each accuracy metric together with a complexity metric. Usually, MoFGBMLML solves each MOP independently. In this paper, we incorporate the idea of multi-tasking optimization into MoFGBMLML so that multiple MOPs are solved simultaneously. We also propose a new information sharing method to improve the effectiveness of multi-tasking optimization in MoFGBMLML. Our experimental results show that multiple accuracy metrics can be simultaneously optimized through the multi-tasking optimization framework and the proposed information sharing method improves the classification accuracy of fuzzy classifiers obtained by MoFGBMLML.