FedRAHi: Reliability-Aware Hierarchical Collaboration for Federated Graph Foundation Models

Abstract

Federated Graph Foundation Models (FedGFM) offer a decentralized GNN training paradigm that combines the collaborative training of federated graph learning (FGL) with the cross-domain generalization of graph foundation model (GFM). However, existing FedGFM methods still suffer from two key limitations, where (i) single-layer codebook quantization induces a representation bottleneck because the same token candidates are assigned to different nodes, reducing node discriminability and causing semantic confusion, and (ii) domain heterogeneity leads to cross-client semantic inconsistency, which weakens the reliability of server-side collaboration and aggregation. Based on these insights, we propose FedRAHi, a Reliability-Aware Hierarchical collaboration framework for FedGFM. Specifically, on the client side, we propose RQ-GAE, a Residual Quantized Graph Autoencoder. It quantizes nodes into tokens in multiple hierarchical codebooks, enhancing node discriminability and alleviating semantic confusion. These hierarchical codebooks can be further uploaded to the server as symbiotic knowledge to harmonize inter-client collaboration, since the codebooks do not directly expose graph data on each client. On the server side, to build a collaboration bridge across clients, we propose a Client Reliability Estimator (CRE). It leverages the symbiotic knowledge to construct a client-aware affinity graph, and performs personalized weighting of client parameters based on the reliability scores. Extensive experiments on seven public graph datasets across four domains and three downstream tasks show that FedRAHi consistently outperforms twelve SOTA baselines from both FGL and FedGFM.