Capped norm based discriminant robust regression learning

Ning Liu

doi:10.1016/j.patcog.2024.111288

Return

Capped norm based discriminant robust regression learning

2025-05-01

PRE

DOI：10.1016/j.patcog.2024.111288

Original

Original request for help

Save

Abstract

En 中文

Exploring underlying correlation structures among data, improving robustness to noise, obtaining discriminant projections, integrating local information of data, and selecting significant features are important objectives in feature extraction methods. To address these objectives, a novel capped norm based discriminant robust regression (CNDRR) method is proposed in this paper. In CNDRR, underlying correlation structures are explored. Besides, the robustness to noise is guaranteed since robust capped norm are used as basic measurement. Moreover, to obtain more discriminant projections and improve the robustness, CNDRR redefines the scatter matrices with a robust L 21-norm based on the differential criterion, discards the square operation on the Euclidean distance in the locality preserving criterion, and employs a robust capped norm as a basic measurement. In addition, by imposing the L 21-norm regularization term on the projection matrix, the joint sparsity of the projection can be enhanced, and the most significant features can be selected. Theoretical analyses, including convergence analyses and computational complexity, are presented. The experimental results show that the proposed CNDRR is superior to the other compared methods on eight datasets.

Keywords:

Capped norm

Discriminant robust regression

Joint sparse subspace learning