Sparse linear discriminant analysis for supervised feature selection

Jianyong Zhu

doi:10.1016/j.neucom.2026.134903

Sparse linear discriminant analysis for supervised feature selection

2026-08-25

0

PRE

AI

DOI：10.1016/j.neucom.2026.134903

Original

Original request for help

Save

Abstract

En 中文

Sparse learning-based feature selection methods have garnered considerable attention in the field of machine learning. In this study, we introduce a novel Linear Discriminant Analysis (LDA) sparsity-based feature selection method, employing a trace difference LDA to effectively select features. The projection matrix is constrained using the l 2 , 0 -norm, promoting a higher level of sparsity compared to the l 2 , 1 -norm. However, solving feature selection problem under the l 2 , 0 -norm constraint poses significant challenges. We address these by decomposing the projection matrix into two components: a discrete selection matrix and a weight matrix. This decomposition transforms the complex optimization problem into a more tractable problem centered on the discrete selection matrix. We employ coordinate descent to solve the l 2 , 0 -norm constrained optimization problem. The methodology proposed herein extends to general optimization issues related to the l 2 , 0 -norm constraint. Extensive experimental evaluations affirm the robust and consistent performance of our method. When compared with alternative approaches, our method demonstrates superior performance, indicating its efficacy in selecting more discriminative features for classification tasks.

Sparse linear discriminant analysis for supervised feature selection

Abstract

Journal

Organization