FedPKA: Federated Graph-Level Clustering Network with Personalized Knowledge Aggregation

Junlong Wu, Haotian Wang, Jingxin Liu, Wenxuan Tu, Renda Han, Jieren Cheng, Xiangyan Tang

Abstract: With the explosive growth of real-world data, the demand for federated graph clustering methods has become urgent. Recently, several solutions have been proposed, generate global consensus through multi-source clustering structures to facilitate local training. However, existing methods fail to incorporate prior confidence, resulting in data distribution discrepancies and knowledge drift, which limit the server’s ability to guide the client. To address this, we propose a novel unsupervised graph learning framework, Federated Personalized Knowledge Aggregation (FedPKA). Specifically, FedPKA assesses the confidence of local domain knowledge structures relative to the global model and employs a dynamic community-aware aggregation model to tailor personalized parameters for each client. This approach effectively handles heterogeneity from non-independent and identically distributed (non-IID) graph data distributions and knowledge drift. Furthermore, we propose an adaptive prototype adjustment mechanism to align more precisely with true prototypes, reducing uncertainty caused by local data drift. Experimental results on 15 cross-dataset and cross-domain non-IID graph datasets demonstrate that FedPKA outperforms existing methods.

External IDs:dblp:conf/icic/WuWLTHCT25