Twin Contrastive Learning for Multi-View Clustering

Multi-view Clustering (MVC) utilizes the complementary and consistent information between multiple views or modal data to more comprehensively reveal the inherent structure and patterns of data. However, mining the invariance and complementarity of multi-view remains a challenging problem. In this paper, we propose to perform online clustering by conducting multi-viewtwin contrastive learning at both the instance-level and cluster-level, extracting multi-level target information from multi-view data in different feature spaces without feature fusion. This includes mixed low-level features reconstructed by autoencoders, instance-level consistency information in twin contrastive learning, and cluster-level discriminative private information. This approach avoids conflicts between learning consistent public semantics and discriminative private information. By selecting high-confidence samples as pseudo-labels and eliminating inherent ‘‘false negative sample pairs’’ in contrastive learning to correct over-clustering, we further improve clustering performance. After the model training converges, it can independently assign clustering to each instance in an end-to-end manner, thus easily adapting to streaming data and online scenarios. Extensive experiments on public datasets demonstrate that our method achieves state-of-the-art clustering effectiveness.