Feature Selection Integrating Decision Importance Grouping and Particle Swarm Optimization

Gene expression data typically exhibits the characteristics of high-dimensionality coupled with a small sample size, along with notable redundancy and high inter-feature correlations. These traits pose significant challenges for gene selection tasks. To tackle this complex issue, this study proposes a novel feature selection algorithm that synergistically combines decision importance subset grouping and particle swarm optimization (DIGPSO). The DIGPSO follows a two-stage processing pipeline. In the first stage, four filter feature selection techniques, namely Pearson correlation coefficient, t-test, Relief algorithm, and information gain, are utilized to identify and rank the most relevant features. Based on the frequency at which each feature is selected across these four methods, the selected features are aggregated into two distinct subsets, which are considered to represent different levels of importance. In the second stage, particle swarm optimization (PSO) is applied to these two subsets respectively to further refine feature selection. After this step, the features selected from the two subsets are combined into a single set. The combined feature set undergoes PSO to select the features with the maximum information volume and non-redundancy. This paper comprehensively assesses the performance of the DIGPSO algorithm and conducts systematic comparative experiments on nine high-dimensional and small-sample omics datasets. The experimental results show that: Compared with the four classical filtering feature selection algorithms, DIGPSO has the best average classification accuracy. DIGPSO exhibits excellent feature compression performance. The classification performance of the selected feature scale is reduced by 48.96% compared with the suboptimal algorithm Relief, while the classification accuracy is increased by 4.9%. Compared with six mainstream algorithms, it is found that DIGPSO can achieve the optimal classification performance on most datasets and show good generalization ability. These experimental results fully verify the effectiveness and superiority of the DIGPSO algorithm in the feature selection task of high-dimensional and small-sample omics data.