An end-to-end Graph Convolutional Network for Semi-supervised Subspace Clustering via label self-expressiveness

• Published in: Knowledge-based systems Vol. 286; p. 111393
• Main Authors: Qi, Tingting, Feng, Xiangchu, Gao, Bian, Wang, Kun
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 28.02.2024
• Subjects: Graph Convolutional Network (GCN); Label self-expressiveness; Semi-supervised learning; Subspace clustering; Graph Convolutional Network (GCN); Subspace clustering; Label self-expressiveness; Semi-supervised learning
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2024.111393

Semi-supervised clustering provides accurate assignments by leveraging a limited amount of supervisory information. There are two main groups of conventional semi-supervised subspace clustering algorithms: Pairwise Constraints (PC)-based methods and Gaussian Fields and Harmonic Function (GFHF)-based methods. Nevertheless, PC-based methods are limited because of their reliance on weak supervisory information. GFHF-based methods, on the other hand, focus on the local relationship between the affinity matrix and the assignment matrix, which restricts their clustering efficacy. Recently, graph convolutional network (GCN) has been employed to effectively transmit the local graph topology of the raw data into the feature domain, which motivates the proposed GCN model for Semi-supervised Subspace Clustering via label self-expressiveness (GSSC). Conventional semi-supervised clustering has been enhanced by incorporating a novel concept known as label self-expressiveness, which serves as a guiding factor for the clustering process. Furthermore, GSSC considers the global geometric structure of the assignment matrix and generates a discriminable assignment matrix using an end-to-end mechanism. Additionally, a well-designed loss function is introduced to train the novel GSSC. The experimental results conducted on four benchmark datasets demonstrate that the proposed GSSC algorithm exhibits superior performance compared to existing semi-supervised clustering algorithms. •A concept of label self-expressiveness which captures the global geometric structure of the assignment matrix sufficiently is proposed.•A novel GCN based model for mapping raw data to an assignment matrix is demonstrated.•A well-designed loss function is presented to effectively train GSSC.