Self-weighted subspace clustering via adaptive rank constrained graph embedding

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 1
• Main Authors: Jiang, Kun, Yang, Zhihai, Sun, Qindong
• Format: Journal Article
• Language: English
• Published: London Springer London 01.03.2025; Springer Nature B.V
• Subjects: Adaptive sampling; Algorithms; Clustering; Computer Science; Constraints; Embedding; Graphical representations; Original Article; Pattern Recognition; Subspace methods; Subspaces; Rank constrained structure; Subspace clustering; Profile-based graph embedding; Self-weighted feature learning
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-024-01405-6

Recent years, subspace clustering methods have attracted wide attention in partitioning high-dimensional data from a union of underlying subspaces, in which the data distribution is mainly explored to compensate for the absence of label information. However, for practical applications, subspace clustering still suffers from redundant and noisy features, which brings about disturbed reconstruction loss and restricts trustworthy graph learning. In this paper, we propose a robust subspace clustering framework via Self-weighted feature learning and adaptive rank constrained graph embedding (SWARG) to address the limitations of existing graph-based subspace clustering models. Specifically, a feature self-weighted learning term is introduced to the sparse subspace clustering framework to alleviate the disturbed contributions from the noisy and redundant features. As such, a few discriminative features will act as remarkable contributions in representing data samples. Meanwhile, the profile-based graph embedding term further preserving the contribution behavior information of data samples that distributed around the same subspace. Moreover, the adaptive rank-constraint graph embedding method is considered to guarantee discriminative structure for different components of representation matrix with flexible entropy-based similarity preserving. To solve the proposed model, we then develop an efficient alternative direction updating algorithm, together with convergence and complexity analysis. Finally, experimental results on toy databases and benchmark databases demonstrate the effectiveness of the proposed SWARG model compared to a series of state-of-the-art models. Our code is available at http://github.com/ty-kj/SAWRG .