Contrastive-Driven Diversity and Consistency Exploration in Tensorized Multiview Subspace Clustering
- 7 November 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 37 (3) , 1492-1504
Abstract
The success of multiview subspace clustering (MVSC) lies in the efficient integration of consensus and complementary information from subspace structures. However, existing MVSC algorithms often treat these two types of information separately, overlooking the obvious correlations—both positive and negative—that exist between them. To address this issue, we propose a novel method called contrastive-driven diversity and consistency exploration in tensorized MVSC (CD-TMSC). Specifically, our method begins by segmenting self-representations into a consensus representation and a set of specific representations to accurately model both consensus and complementary information. Drawing inspiration from contrastive learning, we introduce a novel fractional regularization term to harness both the positive and negative correlations inherent in consensus and complementary information, where the numerator, employing the Hilbert–Schmidt independence criterion (HSIC), quantifies the negative correlation between the consensus and view-specific representations, as well as among the view-specific representations themselves. Conversely, the denominator, also leveraging HSIC, measures the positive correlation between the original data matrices and their self-representations. Minimizing this term has a dual effect: it reduces the numerator, thereby amplifying the negative correlation, a move that might be seen as counterproductive but, in our innovative approach, it promotes the diversity within representation matrices. Simultaneously, it increases the denominator, reinforcing the positive correlation and bolstering the consistency of the information. Additionally, we incorporate a graph regularization term for the consensus matrix to capture more consistent manifold information. Finally, utilizing the high-quality consensus and view-specific representations derived from our constraints, we reconstruct self-representation matrices and construct a third-order tensor with a low-rank constraint to explore higher order correlations within the self-representations. Our method integrates contrastive-driven regularization, manifold learning, and low-rank tensor learning into a cohesive framework, optimized using an alternating direction minimization strategy. Experimental results on multiple benchmark datasets show that our approach outperforms several state-of-the-art methods.Keywords
Funding Information
- National Natural Science Foundation of China (61976005)
- University Natural Science Research Project of Anhui Province (2022AH050970)
- Discipline (Major) Leader Development Project (DTR2025018)
This publication has 40 references indexed in Scilit:
- Tensorized diversity and consistency with Laplacian manifold for multi-view clusteringInformation Sciences, 2024
- A hybrid multilinear-linear subspace learning approach for enhanced person re-identification in camera networksExpert Systems with Applications, 2024
- Integrated Heterogeneous Graph Attention Network for Incomplete Multi-modal ClusteringInternational Journal of Computer Vision, 2024
- Adaptive multi-view subspace learning based on distributed optimizationInternet of Things, 2024
- Deep cross-modal subspace clustering with Contrastive Neighbour EmbeddingNeurocomputing, 2024
- Double-level View-correlation Multi-view Subspace ClusteringKnowledge-Based Systems, 2023
- Seeking commonness and inconsistencies: A jointly smoothed approach to multi-view subspace clusteringInformation Fusion, 2022
- Consistent and diverse multi-View subspace clustering with structure constraintPattern Recognition, 2022
- Constrained bilinear factorization multi-view subspace clusteringKnowledge-Based Systems, 2020
- On Unifying Multi-view Self-Representations for Clustering by Tensor Multi-rank MinimizationInternational Journal of Computer Vision, 2018