Subtopology-Assisted Federated Graph Learning With Adaptive Neighbor Generation in Edge-Client Collaborative Networks

Federated Graph Learning (FGL) facilitates cooperative training on distributed clients for semi-supervised classification tasks. However, many existing FGL frameworks overlook the missing inter-client topology structure in real-world scenarios, leading to inadequate feature aggregation across multi-order neighbor clients. Meanwhile, they commonly adopt the topology-irrelevant FedAvg that neglects the impact of uneven information flow from the distinct subtopologies, resulting in unsatisfactory client performance. Besides, excessive training costs happen as the number of clients grows, causing single-point overload. To address these important challenges, we propose SpreadFGL, a novel FGL framework, to promote the unbiased information flow with collaboration of edge server and clients, extracting more generalized potential relationships between subgraphs based on topology-aware patterns. In SpreadFGL, we design an adaptive graph imputation generator combined with a versatile assessor to explore potential links between multiple clients. Next, a new negative sampling mechanism is proposed to enable SpreadFGL concentrate on more refined information for downstream tasks. Then, a subtopology-assisted aggregation mechanism is proposed for SpreadFGL to mitigate the imbalanced impact of sub-structural patterns. Furthermore, to improve load balancing at the edge layer, SpreadFGL follows distributed training to ensure fast convergence. Extensive experiments on real-world testbed and benchmark graph datasets demonstrate the superiority of the proposed SpreadFGL, which outperforms state-of-the-art algorithms in terms of both classification performance and convergence speed across complex scenarios.