Self-Weighted Supervised Discriminative Feature Selection via Redundancy Minimization

Abstract

Feature selection plays a key role in many machine learning problems. Especially as an important data preprocessing method, robust and pragmatic feature selection methods can be applied to extract meaningful features and eliminate redundant ones. As we all known, many feature selection methods select features by using some certain feature evaluation criteria to obtain the corresponding score for every feature, such that we can select high score features. Unfortunately, correlated features usually connect with each other, which may result in large correlations between top ranked features, such that the redundancy among the selected features is brought about. To solve this problem, we introduce the redundancy matrix A in the AGRM (a novel auto-weighted feature selection framework via global redundancy minimization) framework. Meanwhile, we introduce the adaptive redundancy matrix S and treat the redundancy matrix S as an optimizing variable, rather than setting the redundant matrix S as a prior. In addition, we propose a robust algorithm to efficiently address the constrained optimization problem. Finally, extensive experiments on six datasets show the superiority of our proposed method.

Keywords

• Feature selection
• linear discriminant analysis (LDA)
• redundancy minimization
• sparse regularization