Robust feature selection with minimum redundancy for binary gene expression classification in machine learning

Abstract

High-dimensional gene expression data often contain noisy, irrelevant, and redundant features that can reduce classification accuracy, increase computational burden, and weaken model stability. Therefore, effective feature selection is essential for improving predictive performance and interpretability in biomedical machine learning. This paper proposes a robust feature selection framework with minimum redundancy, called Robust Feature Selection with Minimum Redundancy, for binary classification of gene expression data. The proposed method first constructs a robust class-location estimate for each gene by combining trimmed means, winsorized means, and tri-means. It then computes a normalized robust discriminative score to rank genes by class-separation ability while accounting for within-class variability. To obtain a compact, informative subset of genes, Robust Feature Selection with Minimum Redundancy applies a two-stage redundancy-control strategy: Spearman rank correlation reduces strong monotonic dependencies, while normalized mutual information controls nonlinear dependence among selected genes. We evaluate the proposed method on several publicly available microarray datasets using random forest, support vector machine, and k-nearest-neighbor classifiers. The experimental results show that RFS-MR selects a smaller and less redundant set of informative genes and achieves competitive or improved classification performance compared with existing feature selection methods. These findings suggest that the proposed framework is a robust, interpretable, and redundancy-aware approach for high-dimensional gene expression classification.