Semi-supervised feature selection via adaptive structure learning and constrained graph learning

Abstract

Graph-based sparse feature selection plays an important role in semi-supervised feature selection, which greatly improves the performance of feature selection. However, most existing semi-supervised methods based on graph are still limited in two main aspects. On the one hand, the quality of the similarity matrix will affect the performance of the learning model. Adaptive graph learning improves the quality of similarity matrix by learning the similarity matrix adaptively. However, most methods based on adaptive graph learning ignore the label information, which may limit the quality of the similarity matrix. On the other hand, many state-of-the-art methods only consider the local structure and neglect the global structure of samples, which will result in high redundancy in the selected features. To alleviate the impact of the above problems, in this study, a novel semi-supervised feature selection model named ASLCGLFS is proposed. In the proposed method, the adaptive graph learning is extended through label information, which aims to further improve the quality of the similarity matrix by utilizing the label information to constrain the graph learning. Moreover, adaptive structure learning is introduced, which not only considers the global structure but also facilitates feature selection. An iteration method is designed to solve the objective function and the convergence of this method is proved theoretically and experimentally. Extensive experiments conducted on common datasets verify that the proposed ASLCGLFS is better than some state-of-the-art feature selection algorithms in performance.