Unsupervised feature selection via adaptive graph and dependency score. (July 2022)

• Record Type: Journal Article
• Authors: Huang, Pei; Yang, Xiaowei
• Is Part Of: Pattern recognition. Volume 127(2022)
• Publication Date: 2022-07
• Subjects: Unsupervised feature selection -- Adaptive graph -- Mutual information -- Entropy; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108622
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel unsupervised feature selection method based on adaptive graph learning and dependency score (AGDS) is proposed. AGDS can make each sample adaptively learn from its k nearest neighbors so that the problem about imbalanced neighbors and trivial solution can be prevented. AGDS introduce dependency score by incorporating mutual information and entropy to measure feature uncertainty and feature pairwise dependency so that more redundant features can be eliminated. Extensive experiments conducted on 13 benchmark datasets show the effectiveness of AGDS. Abstract: Unsupervised feature selection is an important topic in the fields of machine learning, pattern recognition and data mining. The representation methods include adaptive-graph-based methods and self-representation-based methods. The former methods have a longstanding and undiscovered problem about imbalanced neighbors, and the latter ones do not perform well when features are not linearly dependent. To deal with these problems, a novel unsupervised feature selection method is proposed to ensure k connectivity and eliminate more redundant features based on adaptive graph and dependency score (AGDS). Extensive experiments conducted on 13 benchmark datasets show the effectiveness of AGDS.