A High-Dimensional Feature Selection Algorithm Based on Multiobjective Differential Evolution
Zhenxing Zhang, Qianxiang An, Yilei Wang, Chenfeng Wu, Baoling Dong, Chunjie Zhou
TL;DR
This work tackles high-dimensional multiobjective feature selection by minimizing feature count and classification error using MODE-FS, a differential-evolution-based framework enhanced with FCM-derived feature weights and a cosine-based redundancy index. The method introduces three key components—WRBI for diverse initialization, MSBIU for redundancy-aware mutation, and FOAGM for adaptive grid-based Pareto-front refinement—driven by the feature weight matrix $Q$, redundancy index $A$, and threshold $ au$. Empirical results on 11 UCI datasets show MODE-FS consistently outperforms five strong MOEA baselines in hypervolume and IGD, yielding well-distributed and near-optimal Pareto fronts. The approach demonstrates robust performance in high-dimensional settings and offers practical benefits for efficient, accurate feature subset selection in complex data environments.
Abstract
Multiobjective feature selection seeks to determine the most discriminative feature subset by simultaneously optimizing two conflicting objectives: minimizing the number of selected features and the classification error rate. The goal is to enhance the model's predictive performance and computational efficiency. However, feature redundancy and interdependence in high-dimensional data present considerable obstacles to the search efficiency of optimization algorithms and the quality of the resulting solutions. To tackle these issues, we propose a high-dimensional feature selection algorithm based on multiobjective differential evolution. First, a population initialization strategy is designed by integrating feature weights and redundancy indices, where the population is divided into four subpopulations to improve the diversity and uniformity of the initial population. Then, a multiobjective selection mechanism is developed, in which feature weights guide the mutation process. The solution quality is further enhanced through nondominated sorting, with preference given to solutions with lower classification error, effectively balancing global exploration and local exploitation. Finally, an adaptive grid mechanism is applied in the objective space to identify densely populated regions and detect duplicated solutions. Experimental results on 11 UCI datasets of varying difficulty demonstrate that the proposed method significantly outperforms several state-of-the-art multiobjective feature selection approaches regarding feature selection performance.