Uncertainty-Aware Federated Graph-Level Clustering with Missing Attributes

Federated graph-level clustering (FGC) has attracted increasing attention for privacy-preserving multi-graph analysis. However, existing methods typically assume fully observed node attributes, which often fails in cross-domain non-independent and identically distributed (non-IID) settings due to device failures or limited data access, while FGC with missing attributes remains largely underexplored. To address this, we propose **U**ncertainty-Aware **Fed**erated **G**raph-Level **C**lustering with Missing Attributes **(UFedGC)**, which couples attribute completion and graph-level clustering through uncertainty-aware closed-loop learning and cross-client semantic alignment. Specifically, UFedGC learns disentangled shared and private graph representations and uses evolving cluster assignments to guide attribute completion; the completed graphs are re-encoded to refine the assignments, forming a bidirectional interaction between attribute inference and clustering. Local clusters are then summarized into uncertainty-aware distributional statistics for reliable global semantic alignment and knowledge aggregation across heterogeneous clients, so that uncertain or domain-specific information contributes adaptively rather than being blindly incorporated. Extensive experiments demonstrate the superiority of UFedGC.