Multi-view clustering is an important approach to mining the valuable information within multi-view data. In this paper, we propose a novel multi-view deep subspace clustering method based on contrastive learning and Cauchy-Schwarz (CS) divergence. Our method not only uses contrastive learning techniques and block diagonalization constraints to guide representation matrix learning, but also combines representation learning and clustering processes to achieve the interaction of representation and clustering. First, we introduce a novel loss function based on CS divergence in the clustering module to achieve the interaction of representation and clustering. Second, we propose an extension of the multiple positive and negative pair diffusion method to enhance contrastive learning. Finally, we establish the equivalence between contrastive clustering and spectral clustering with orthogonal constraints, leading to a comprehensive model optimization. We evaluate our method on six publicly available datasets and compare its performance with eight competing methods. The results demonstrate the superiority of our method over the compared multi-view clustering methods.

Incomplete multi-view clustering focus on mining useful information from low-quality multiple sources, such as missing and distorted data that are prevalent in real life. However, after representation learning and the processing of incomplete information, existing methods often leave representations containing information task-irrelevant information. In addition, the separation between missing data imputation and clustering tasks leads to sub-optimal multi-view clustering performance. To address these issues, we propose an incomplete multi-view clustering method based on mutual information. For the problem of task-irrelevant information, we use incomplete view prediction to extract sufficient and minimal task-relevant information and provide theoretical proof from the perspective of mutual information. For the problem of separation between missing data imputation and clustering tasks, we integrate incomplete-view prediction with contrastive clustering, collaboratively enhancing the clustering performance. Comparative experiments on five public datasets, under both complete and incomplete scenarios, reveal that our method outperforms nine other competing approaches, demonstrating its effectiveness and robustness in handling multi-view data.