Adaptive graph learning and low-rank constraint for supervised spectral feature selection

摘要

Spectral feature selection (SFS) effectively improves performance of feature selection by introducing a graph matrix to preserve information of data. However, conventional SFS (1) generally preserves either global structure or local structure of data in selected subset, which is not capable of providing comprehensive information for model to output a robust resu (2) constructs graph matrix by original data, which usually lead to a suboptimal graph matrix because of redundant information; (3) conducts feature selection task depending on the fixed graph matrix, which is easily trapped in local optimization. Thus, we have proposed a novel SFS to (1) preserve both local information and global information of original data in feature-selected subset to provide comprehensive information for learning model; (2) integrate graph construction and feature selection to propose a robust spectral feature selection easily obtaining global optimization of feature selection. Besides, for the proposed problem, we further provide a optimization algorithm to effectively tackle the problem with a fast convergence. The extensive experimental results showed that our proposed method outperforms state-of-the-art feature selection methods, in terms of classification performance.