Multi-view clustering via contrastive approach

Existing traditional graph-based multi-view clustering models construct graphs independently for each view, neglecting cross-view information interaction and thereby limits the clustering performance. To address this problem, we introduce the concept of contrastive learning into traditional multi-view clustering, and propose Multi-view Clustering via Contrastive Approach (MVCCA) model, which is designed based on the principles of alignment and uniformity. Specifically, the consensus block diagonal graph can be learned to mine positive sample pairs, which facilitates information interaction across different views and lays a solid foundation for the alignment of cross-view representations. For uniformity, representation variance maximization is employed to ensure that all sample representations are evenly distributed in the latent space, which increases the discriminative power of the learned representations. Subsequently, a bidirectional fusion strategy is employed to construct the model, which enables it to adaptively learn to balance alignment and uniformity. Finally, an optimization algorithm is developed to effectively optimize the MVCCA model. Extensive experiments validate the feasibility and effectiveness of the proposed model.