Research map / Graph learning / Graph self-supervised learning

Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FedATH, a federated graph learning method that uses a learned edge evaluator to split each client's graph into causal and biased subgraphs, trains dual GNNs with CE, negative-entropy and HSIC losses, and shares only the causal GNN.

Problem

Federated graph learning suffers from topology heterogeneity: nodes of the same class connect to different classes across clients, causing biased local GNN training; existing fixes add global signals at high computation and communication cost.

Approach

Instead of adding global information, reduce superfluous topology: a private edge evaluator splits local edges into causal and biased subgraphs, a dual-GNN encodes each, cross-entropy and negative-entropy losses plus HSIC enforce separation, and only the causal GNN is aggregated.

Key results

• On Cora with 10 clients, FedATH reaches 77.90% ACC versus 74.47% for the best baseline FGSSL.
• On Roman-empire with 10 clients, FedATH reaches 48.18% ACC versus 41.59% for FedSage+.
• On ogbn-arxiv with 10 clients, FedATH reaches 42.21% ACC versus 41.02% for FedSage+.

Measured effects reported in the paper

Surprising findings

• Several conventional FL methods (e.g., FedProx, MOON) perform no better than or worse than FedAvg in many settings, indicating they do not handle topology heterogeneity.
• Sharing the edge evaluator or biased GNN in addition to the causal GNN degrades performance.

Limitations

• Requires node labels for training, so it is restricted to supervised settings as the authors note.
• The lambda trade-off parameter must be tuned per dataset (larger on Cora, smaller on WikiCS).
• Theoretical analysis is a generalization bound whose link to the method is only qualitative.

Experimental setup

Datasets: Cora PubMed ogbn-arxiv Photo WikiCS Roman-empire

Benchmarks: Cora PubMed ogbn-arxiv Photo WikiCS Roman-empire

Metrics: ACC F1

Baselines: FedAvg FedProx MOON FedOPT FedProto FedSage+ FGSSL FedPUB FedTAD

Builds on: FedAvg FedTAD FGSSL FedPUB

Closest papers by meaning

• FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.92)
• $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.91)
• FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.90)
• GHOST: Generalizable One-Shot Federated Graph Learning with Proxy-Based Topology Knowledge Retention (ICML 2025 · similarity 0.90)
• Virtual Nodes Can Help: Tackling Distribution Shifts in Federated Graph Learning (AAAI 2025 · similarity 0.90)
• FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.89)