Deep contrastive coordinated multi-view consistency clustering

Fuhao Shi, Shaohua Wan, Shengli Wu, Hui Wei, Hu Lu

Abstract: Multi-view clustering (MVC) aimed at partitioning data samples into coherent clusters by integrating information from multiple perspectives. Recently, deep contrastive learning approaches have exhibited substantial capabilities in feature extraction within MVC frameworks. However, the challenge lies in extracting efficient feature representations while ensuring consistency. Moreover, existing deep clustering methods based on contrastive learning often overlook the consistency of cluster representation during clustering processes. In this study, we address these challenges by proposing a novel deep learning method called Contrastive Coordinated Multi-View consistency Clustering (CCMVC). Our approach leverages contrastive learning to coordinate training across three levels: feature, cluster, and view. Specifically, we enhance clustering performance by implementing an alignment method to ensure consistent information alignment across different views. This method assigns semantically similar representations for clustering tasks and effectively explores shared semantics across views while mitigating view-specific noise. Experimental evaluations conducted on seven datasets demonstrate the efficacy and superiority of our proposed CCMVC method over existing state-of-the-art approaches. Code is available at https://github.com/hulu88/CCMVC.