TopFGL: A Topology-Aware and Distributionagnostic Federated Learning Framework Tackling Topological Heterogeneity on Graph Data

While the modern internet generates graph data at an unprecedented scale, stringent privacy regulations like GDPR have fragmented it into silos, creating an urgent need for distributed learning paradigms. Federated Graph Learning (FGL) meets this demand, enabling collaborative training across graph silos. A core challenge for FGL is that the graphs among participants are topologically heterogeneous due to diverse sources and methods of structure collection, which leads to significant performance degradation. However, recent works tackling heterogeneity require clients to share class-wise intermediate information derived from local embeddings, which could increase risks of label and structure leakage. Moreover, they are often computationally expensive and narrowly tailored for a single type of heterogeneity, which hinders their applicability in efficiency-demanding and distribution-variant practical scenarios. To this end, we present TopFGL, the first topologyaware and distribution-agnostic framework to tackle heterogeneity without intermediate information sharing. TopFGL trains a topology learner in each client to learn local topological patterns, and globally aggregates all learners along with the main task models. We propose a multi-level topology extraction scheme to adapt to diverse distributions, along with a streamlined training pipeline that reduces redundant computations. Building upon the learners, we propose a server-side topological similarity-based aggregation algorithm to share similar optimization directions across clients. A client-side dual-model guided topology augmentation approach is also proposed to supplement neighborhood connections for topology-insufficient nodes, improving adaptability to imbalanced distributions. Experiments on 11 datasets with 3 topological distributions across 3 100 client counts show that TopFGL achieves up to 5.6% accuracy improvement along with 50% time cost reduction compared to state-of-the-art baselines.