Federated incomplete multi-view graph clustering via view completion

Multi-view clustering is a fundamental unsupervised learning paradigm designed to partition multi-view data into disjoint clusters. In real-world scenarios, however, multi-view data are often distributed across various devices, institutions, or edge nodes, motivating the development of federated multi-view clustering. Within this decentralized setting, issues such as missing views, privacy constraints, and high communication overhead pose significant challenges. To address these limitations, we propose a novel framework termed Federated Incomplete Multi-View Graph Clustering via View Completion (FMVC-VC). Specifically, FMVC-VC seamlessly integrates client-side missing-view recovery with server-side global structure optimization. By leveraging view-specific anchor graphs as lightweight knowledge carriers, the framework avoids direct feature-level exposure while substantially reducing communication costs. At the client level, global structural information from the server guides the reliable imputation of missing-view features. Concurrently, at the server level, we introduce a contrastive learning mechanism to refine cross-client structural consistency and further enhance the discriminative capacity of the global anchor graph. Extensive experiments on multiple benchmark datasets demonstrate that FMVC-VC significantly outperforms current federated multi-view clustering methods, particularly in challenging incomplete-view scenarios.