Multi-Layer Contrastive Multi-View Clustering via Spectral Embedding Similarity

Multi-view clustering, which partitions data samples into different categories in an unsupervised manner by learning a consensus representation, has attracted widespread attention. In recent years, contrastive learning has demonstrated great potential in multi-view clustering due to its powerful cross-view feature alignment capabilities. However, existing contrastive multi-view clustering methods suffer from false negative and false positive sample pair issues, which limit further improvement in clustering performance. To address these problems, this paper proposes a Multi-layer Contrastive Multi-view Clustering method Based on Spectral Embedding Similarity (MCSES). By introducing a multi-layer contrastive learning mechanism based on spectral embedding similarity, the proposed method effectively alleviates the performance bottleneck caused by inaccurate sample pair division. Specifically, MCSES first constructs a Laplacian matrix based on the latent representation of each view and utilizes its eigenvectors to generate a spectral embedding similarity matrix. Then, MCSES constructs positive and negative sample pairs based on the global prior information provided by the spectral embedding similarity matrix, thereby mitigating the interference of sample division errors on latent representation learning. Meanwhile, MCSES designs a multi-layer contrastive learning framework. At the cross-view level, a cross-view contrastive loss is designed to fully exploit the complementary information among views. In addition, MCSES adaptively fuses the latent representations of all views with learnable weights to obtain a global multi-view representation, based on which a global contrastive loss is designed to emphasize the consistency information among multiple views. Extensive experiments on public multi-view datasets verify the effectiveness of the proposed method.