FedUGCN: Probabilistic Federated Clustering with Unknown Global Cluster Number

Federated clustering aims to uncover global clustering structures from distributed unlabeled data without centralizing raw samples. Existing methods typically assume that the global number of clusters is known and recover global clustering structures by aggregating client-specific cluster centers. However, the global cluster number is often unknown in practical settings, while heterogeneous client distributions obscure the global clustering structure and complicate the identification of corresponding clusters across clients. To address this issue, we propose a probabilistic $\underline{\text{Fed}}$erated clustering framework for $\underline{\text{U}}$nknown $\underline{\text{G}}$lobal $\underline{\text{C}}$luster $\underline{\text{N}}$umbers ($\text{FedUGCN}$), which jointly infers the number of global clusters and identifies the corresponding global clustering structure through hierarchical aggregation of local clusters across clients and federated probabilistic modeling. Specifically, each client first summarizes its local clusters with cluster-level statistics. The server then employs the Bayesian information criterion to assess evidence for merging clusters across clients and applies constrained hierarchical aggregation to construct candidate global structures. For each candidate, federated probabilistic optimization estimates shared global cluster distributions and client-specific cluster compositions, refining cross-client cluster correspondences while preserving local heterogeneity. Finally, the integrated completed likelihood criterion is designed to select the global number of clusters and latent clustering structure. Extensive experiments demonstrate that FedUGCN achieves competitive clustering performance without requiring prior knowledge of the global cluster number.