Neighborhood Pseudo-Graph Based Personalized Federated Graph Learning

As mobile devices continue to evolve, massive amounts of graph data are collected and stored in them. However, training them jointly faces huge challenges due to extreme bandwidth limitations through wireless communications. Federated graph learning (FGL) has attracted significant attention as a promising solution to cooperative learning over graph data stored in distributed systems. Current FGL methods mainly address two issues through aggregating local linking patterns: graph heterogeneity and the absence of cross-client neighbor information. However, such methods may bring generalized global patterns that introduce unnecessary interference to local GNN. Meanwhile, we find that supplementing a client's local graph with its neighbor information on the global graph improves local model performance and avoids interference from linking patterns in other clients. To address this issue, we propose FedNeighbor, a federated graph learning framework that compensates for missing cross-client neighbors without aggregating heterogeneous global linking patterns. The core idea is to approximate each client's missing global k-hop neighborhood using a lightweight pseudo-graph. By selectively matching relevant pseudo-nodes across clients, FedNeighbor enhances local representation learning under strict bandwidth constraints. We further provide theoretical analysis of privacy preservation and communication efficiency. Extensive experiments on four benchmark datasets demonstrate state-of-the-art node classification performance with significantly reduced communication cost.