Dual-stage Contrastive Learning-enhanced Multi-view Variational Clustering

Multi-view clustering aims to obtain a consensus clustering by integrating complementary and consistent information from multiple views. However, two critical challenges still exist in variational methods: (1) view heterogeneity and noise often make fusion unreliable; (2) ambiguous posteriors and misassigned boundary samples impact the clustering performance. To address these issues, we propose Dual-stage Contrastive Learning-enhanced Multi-view Variational Clustering (DCL-MVC), which integrates contrastive learning into both the fusion and representation stages. Firstly, at the fusion stage, we introduce a fusion-then-attention mechanism to capture cross-view interactions and learn view-level attention weights for building a unified and reliable fused representation, and further introduce instance-level contrastive learning to enforce cross-view alignment at the instance level. Secondly, we focus on boundary samples with uncertain posteriors and refine their cluster assignments by using cluster-center contrastive loss to enlarge inter-cluster margins, while leveraging prototypical contrastive learning with a confidence-aware curriculum to promote intra-cluster compactness at the representation stage. Extensive experiments on six real-world datasets demonstrate consistent improvements over strong baselines and validate the contribution of each component.

Many real-world datasets describe the same object in several ways: an image may come with text, or a scene may be recorded by multiple sensors. Clustering such data without labels is difficult because different views can be noisy, disagree with each other, or make some examples hard to assign to a single group. Our work introduces DCL-MVC, a method that first learns how much to trust each view and aligns different views of the same example before making a shared representation. It then pays special attention to uncertain “borderline” examples that lie between clusters, encouraging cluster centers to move apart and examples in the same cluster to become more compact. This two-stage design combines a data-generation model with a learning strategy that compares similar and dissimilar examples, so the model can both explain the data and form clearer groups. In experiments on six real-world multi-view datasets, DCL-MVC generally produced more accurate and stable clusters than several recent methods. This can help applications such as image organization, retrieval, classification, and recommendation, where labels are expensive but data often arrive from multiple sources.