Weight matrix sharing for multi-label learning. (April 2023)

• Record Type: Journal Article
• Authors: Qian, Kun; Min, Xue-Yang; Cheng, Yusheng; Min, Fan
• Is Part Of: Pattern recognition. Volume 136(2023)
• Publication Date: 2023-04
• Subjects: Low-rank -- Missing labels -- Multi-label learning -- Shared weight -- Sparse; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.109156
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a shared weight matrix with low-rank and sparse regularization for multi-label learning (2SML) algorithm. The feature manifold and the label manifold share the weight matrix. We use high representativeness instances to learn implicit correlations for sparse labels. We employ nuclear norm to model low-rank structure for missing labels. We employ l 1 norm to learn label-specific feature for sparse structure. Abstract: Multi-label learning on real-world data is a challenging task due to sparse labels, missing labels, and sparse structures. Some existing approaches are effective in addressing the former two issues. In this paper, we propose a shared weight matrix with low-rank and sparse regularization for multi-label learning (2SML) algorithm to address the issues simultaneously. First, two explicit correlation matrices are constructed from the feature matrix and label matrix. Second, we select informative labels by instance representativeness to learn implicit correlations. Third, a feature manifold and label manifold are employed to guide the shared weight learning process. Extensive experiments are undertaken on multiple benchmark datasets with and without missing labels. The results show that the proposed method outperforms the state-of-the-art methods.