Abstract Multi-objective feature selection (MOFS) aims to identify the most discriminative feature subset by simultaneously optimizing two conflicting objectives: minimizing the number of selected features and reducing the classification error rate. However, current MOFS algorithms face several challenges, including uneven population initialization, mutation operators that overlook feature correlations, and crowding in dense regions that reduces solution diversity. These issues collectively compromise search efficiency and the convergence quality of the Pareto front. In this paper, we present an enhanced differential evolution methodology designed to seek multiple optimal feature subsets. First, a balanced diversity initialization strategy is introduced that leverages feature weights and redundancy indices to enhance both diversity and uniformity within the initial population. Subsequently, a mutation strategy informed by these weights and redundancy guides mutations while employing non-dominated sorting to prioritize solutions with lower classification errors, thereby balancing global exploration with local exploitation. Finally, a grid-aware crowding regulation approach is proposed to identify dense areas in objective space and eliminate redundant solutions. Experimental results derived from 13 UCI datasets of varying complexity illustrate that our proposed method significantly outperforms several state-of-the-art MOFS techniques in terms of feature selection efficacy.
Zhang et al. (Wed,) studied this question.