Covariance-driven consistency and MMD complementarity with hybrid graph for multi-view clustering

The key to multi-view clustering lies in effectively learning a high-quality unified representation that simultaneously captures both the consistent structural relationships across views and the complementary information specific to each view. To achieve this objective, this paper proposes a novel multi-stage multi-view clustering framework. First, the local geometric structure of samples is captured at different scales through the concurrent construction of K-nearest neighbours (KNN) graphs and hypergraphs. Subsequently, to generate a more robust unified affinity graph, an Edge-Boost fusion strategy is introduced, which significantly enhances high-confidence connections co-occurring in both topological structures during graph integration. Building upon this unified affinity graph, a Graph Neural Network (GNN) is then employed to perform deep refinement of view features. Furthermore, during the training phase, a complementarity loss based on Maximum Mean Discrepancy (MMD) is integrated to explicitly encourage different views to learn differentiated information. Finally, to overcome the limitations of simple linear averaging, a Covariance-driven Basis Projection (CBP) fusion module is developed. By extracting optimal shared principal directions, this module adaptively aggregates the refined multi-view representations, ultimately forming a highly discriminative unified embedding.