RDFS-TDC: Robust discriminant feature selection based on improved trace difference criterion

Abstract

Various discriminant feature selection models have been proposed that combine discriminant subspaces and sparse constraints. However, most scholars ignore the sensitivity of the discriminant criterion to outliers. In this study, we propose a robust discriminative feature selection method called RDFS-TDC. RDFS-TDC learns the optimal discriminative projection based on the trace-difference criterion, which provides good flexibility while avoiding singular matrices. Subsequently, the objective function was optimized using an iterative reweighting method, which reduced the impact of outliers on the discriminant subspace during the learning process. To satisfy different sparsity requirements, this study introduces the $L_{2,p}$ norm constraint to impose row sparsity on the projection matrix. RDFS-TDC obtained 87.05%, 94.68%, 84.82%, and 89.60% accuracies on YaleB, COIL20, CMUPIE, and FERET, respectively, and the misclassification error rate was 0.01%-3.32% lower compared to other methods. In addition, RDFS-TDC performed better on datasets with different scenarios compared to SDFS, WDFS, Fisher Score, DLSR, ReliefF, and RFS.