Dual-metric guided multi-strategy hybrid optimization for feature selection on high-dimensional medical data

Abstract

The high-dimensional feature selection (FS) problem is challenging in medical fields due to the “curse of dimensionality” and the intricate relationships among various features. Although hybrid FS methods achieve high-performance solutions according to various mutual information metrics, such as symmetrical uncertainty (SU) and maximal information coefficient (MIC), they often overlook the differences between these metrics, and are still need to improve search strategies to escape from local optima. To address these challenges, we propose a dual-metric guided multi-strategy hybrid FS method (DGM) for high-dimensional medical datasets. The importance of features are first evaluated based on the SU and MIC metrics, and then the redundancy between features are reduced by fast clustering and grouping strategies. Furthermore, a two-level sampling strategy is proposed to guarantee the diversity and complementarity of population by considering the Jaccard Similarity and the correlation between features. A novel set-based multi-population PSO is designed to collaboratively search the optimal feature subset while obtaining feature importance during the evaluation process by a tri-archive assisted evolution approach. Specifically, two local archives help individuals escape from local optima, while the global archive optimizes the population. Finally, we develop various squeeze-expand mechanisms to dynamically adjust both the search space and the length of individuals to effectively balance exploration and exploitation. The experimental results on 13 medical datasets show that DGM significantly improves classification performance while selecting fewer features. The T-test results further indicate that DGM significantly outperforms all comparison methods in classification performance on 10 datasets, highlighting its strong competitiveness.