Global and local combined contrastive learning for multi-view clustering

Multi-view representation learning plays a pivotal role in numerous multi-view tasks, including clustering and classification. However, there are two challenging issues that plague these fields: (i) deriving robust multi-view representations from high-quality unlabeled data, and (ii) striking a balance between view consistency and specificity. To address these issues, this paper introduces a novel joint contrastive fusion method aimed at extracting the robust public representations of the input views from unlabeled data. Specifically, our method incorporates an additional representation space and aligns representations within this space, thereby facilitating the learning of robust view public representations. Furthermore, we employ a pseudo-label clustering method to ensure that the model does not converge to a trivial solution. Experiment results demonstrate that our method surpasses 12 competitive multi-view approaches in clustering performance across three real datasets.