Hierarchical feature based dual contrastive multiview clustering

Multiview clustering (MVC) has attracted considerable attention due to its capability to leverage complementary information from multiple views. However, most existing deep MVC approaches primarily focus on learning a unified global representation from each view, which often results in suboptimal exploitation of inter-view complementarity and an inadequate balance between consistency and complementarity. To overcome these limitations, a novel deep MVC framework termed hierarchical feature based dual contrastive learning (HFDCL) is proposed in this paper. Specifically, HFDCL introduces a hierarchical self-attention-based learning architecture that explicitly disentangles the consistency objective from the reconstruction objective. This hierarchical design ensures that the learned representations effectively preserve both localized and holistic features intrinsic to each view. To further enhance the trade-off between consistency and complementarity, HFDCL incorporates a dual contrastive learning module based on information noise contrastive estimation (InfoNCE). This module aims to maximize the mutual information between complementary view-specific representations and the global representation, thereby enabling contrastive fusion across multiple views. This strategy promotes both the diversity and consistency of multiview features, which is crucial for MVC tasks. Extensive experiments conducted on several real-world multiview datasets validate the effectiveness of the proposed HFDCL framework, demonstrating the superior performance in comparison to existing state-of-the-art deep MVC methods.