A Multiform Evolutionary Approach with Enhanced Diversity and Opposition-Based Learning for Multi-Objective Feature Selection

Recently, there has been extensive research on multi-objective feature selection (FS). In general, minimizing the number of selected features and maximizing classification performance are two primary objectives of most multi-objective FS tasks. However, due to complex interactions among features, a subset with poor classification performance does not necessarily imply uselessness; for instance, some features, when combined with others, can significantly enhance classification performance. Therefore, not only the classification performance of feature subsets but also their diversity in the search space should be considered. Motivated by this, this paper proposes an evolutionary multitasking based optimization framework, named multi-objective feature selection with enhanced diversity and opposition-based learning. Firstly, it utilizes two different forms of single-objective FS tasks to provide feature subsets with better classification performance and greater diversity for the multi-objective FS task at hand, thereby promoting evolutionary search in multi-objective FS tasks. Then, to explore more feature combinations, a novel crossover operator utilizing opposition-based learning is proposed to search for more promising feature subsets. Additionally, a duplicate solution handling approach is proposed to maintain population diversity. Experimental results on ten classification datasets demonstrate that the proposed method outperforms five state-of-the-art FS methods in most cases.