Multi-view clustering with semantic fusion and contrastive learning

Abstract

Multi-view data obtained from different sources provides a comprehensive way to model the real world. As one of the fundamental analyses, clustering for multi-view data has attracted increasing attention from researchers in different fields. Although existing multi-view clustering methods have achieved promising results, how to fully utilize the consensus and differences among different views and how to integrate their features to learn a unified representation is still challenging. To overcome the limitations, we put forward a Semantic fusion and Contrastive learning model for Multi-View Clustering (SCMVC) in this paper. SCMVC employs the view-specific encoders to extract embeddings from each view, and reconstructs the topological and attribute information by using the corresponding decoders. Simultaneously, a cross-view contrastive learning module is introduced in SCMVC to assimilate the consensus information across different views. In addition, to distinguish the differences in semantic information across different views, we incorporate an attention mechanism to determine the relative importance of each view in the semantic information fusion. The proposed model is jointly optimized by training and clustering to mutually reinforce each other, gradually improving its performance and robustness. Experiments on three real-world multi-view datasets demonstrate the effectiveness of our SCMVC framework. Source codes are freely available at https://github.com/HexonBang/SCMVC.