Dual contrastive learning with graph masking: A self-supervised framework for multi-view clustering

Graph-based deep multi-view clustering has recently emerged as a prominent research paradigm, driven by its efficacy in modeling nonlinear feature relationships. However, two critical limitations persist in existing approaches: (1) They exhibit high susceptibility to noise and redundancy, which often distorts the graph construction process; and (2) Excessive emphasis on cross-view feature alignment inadvertently neglects inherent inter-view heterogeneity, thereby risking representation homogenization across views. While masking mechanisms can alleviate these issues, the high masking ratio, which is necessary for generating discriminative representations in single-view scenarios, forces graph-based models to produce diverse view-specific representations at the expense of significant consistency loss, i.e., semantically misaligned representations. To address these challenges, this paper proposes Dual Contrastive Masked Graph-Autoencoder Learning (DCMGAL) for multi-view clustering. DCMGAL incorporates a masked aggregation module that amplifies inter-view discrepancies while suppressing noise and redundancy through randomized edge masking. It further designs a global feature fusion mechanism, which integrates dual attention networks with a self-expression network to capture complementary information across views. Moreover, a dual contrastive learning module is devised to enforce cluster-level consistency via cross-view cluster consistency learning and preserve local topological structure via local topology preservation learning, thereby enhancing representation separability. Finally, an adjacency graph reconstruction component is incorporated to retain essential neighborhood information via a graph autoencoder architecture, while simultaneously reconstructing node features. Extensive experiments on benchmark datasets show that DCMGAL significantly outperforms state-of-the-art clustering methods.