Deep multi-view graph clustering with multi-level contrastive guidance

Deep graph-based multi-view clustering has gained prominence for modeling structural relationships and semantic patterns across views. Existing methods typically construct a graph followed by graph neural network processing, but suffer from two key limitations: (1) clustering performance is often constrained by defects in the graph construction; (2) most methods employing contrastive learning prioritize embedding discriminability while neglecting clustering adaptability. To overcome these challenges, we propose the Deep Multi-Level Contrastive Graph Clustering (DMCGC) approach. Specifically, feature-level contrastive learning is introduced to enhance cluster separability within each view and generate clustering-friendly attribute embeddings. Then, dual-scale graph learning is performed to extract view-shared local and global graph embeddings. Finally, attribute and graph embeddings are mapped into soft clustering assignments, regularized by cluster-level contrastive learning for structural consistency. Since the attribute embeddings convey the original characteristics and the graph embeddings are gradually optimized, the dependence on the initial graph is alleviated. Additionally, the clustering adaptability is also improved by the multi-level contrastive learning mechanism. Extensive experimental results on multi-view datasets demonstrate the competitive performance of DMCGC against state-of-the-art algorithms.