Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity
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 FedSPA, a federated graph learning framework that decouples homophilic and heterophilic feature propagation channels (SFPD) and weights client aggregation by homophily bias measured via high-frequency area and Fisher information (HBDA).
Problem
Federated graph learning suffers when clients have different local graph homophily levels (homophily heterogeneity), causing homophily conflict in feature propagation and unequal client contributions to a global GNN.
Approach
Decouple feature propagation into separate channels for homophilic and heterophilic edges (with a harmonizer for edges of unknown label relation) so each channel trains on a unified homophily level across clients, and aggregate client updates with weights that favor clients whose homophily bias (spectral high-frequency area discrepancy for the harmonizer, Fisher information sensitivity discrepancy
Key results
- FedSPA reaches 94.4% on Coauthor-CS vs 91.8% for FedAvg, and 77.0% on Cornell vs 72.3% for FedAvg.
- On Minesweeper, FedSPA achieves 83.2% vs 81.1% for FedAvg; ablation shows SFPD alone and HBDA each add gains over FedAvg (e.g., Arxiv-year 35.6 to 37.26 with SFPD, 38.12 with full HBDA).
- FedSPA dominates in six of seven settings in the main comparison table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Cora | accuracy: +1.7 | Mixed |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · CS (Coauthor-CS) | accuracy: +2.6 | Helped |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Physics (Coauthor-Physics) | accuracy: +1.0 | Helped |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Cornell | accuracy: +4.7 | Helped |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Wisconsin | accuracy: +4.6 | Helped |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Minesweeper | accuracy: +2.1 | Helped |
| use FedSPA instead of FedAvg | ACM-GCN, 2 layers · Arxiv-year | accuracy: +2.5 | Helped |
| add SFPD (ablation) | FedAvg on Coauthor-CS, Minesweeper, and Arxiv-year · Coauthor-CS, Minesweeper, Arxiv-year | accuracy: 91.78 ± 0.31 -> 92.29 ± 0.45; 81.14 ± 0.21 -> 81.62 ± 0.12; 35.60 ± 0.56 -> 37.26 ± 0.88 | Helped |
| add HBDA to SFPD (ablation) | FedSPA on Coauthor-CS, Minesweeper, and Arxiv-year · Coauthor-CS, Minesweeper, Arxiv-year | accuracy: 92.29 ± 0.45 -> 94.37 ± 0.13; 81.62 ± 0.12 -> 83.21 ± 0.21; 37.26 ± 0.88 -> 38.12 ± 0.30 | Helped |
| add prior bias-driven aggregation (ablation) | SFPD on Coauthor-CS, Minesweeper, and Arxiv-year · Coauthor-CS, Minesweeper, Arxiv-year | accuracy: 92.29 ± 0.45 -> 93.31 ± 0.39; 81.62 ± 0.12 -> 82.39 ± 0.42; 37.26 ± 0.88 -> 37.77 ± 0.52 | Helped |
| add posterior bias-driven aggregation (ablation) | SFPD on Coauthor-CS, Minesweeper, and Arxiv-year · Coauthor-CS, Minesweeper, Arxiv-year | accuracy: 92.29 ± 0.45 -> 93.56 ± 0.74; 81.62 ± 0.12 -> 82.67 ± 0.23; 37.26 ± 0.88 -> 37.42 ± 0.68 | Helped |
Limitations
- Evaluation limited to node classification under Louvain-simulated client splits and a single backbone (ACM-GCN).
- Homophily bias proxies (high-frequency area, FIM) are indirect substitutes for the unavailable homophily level due to invisible labels.
Experimental setup
Datasets: Cora Coauthor-CS Coauthor-Physics Cornell Wisconsin Minesweeper Arxiv-year CSBM (synthetic)
Benchmarks: Cora Coauthor-CS Coauthor-Physics Cornell Wisconsin Minesweeper Arxiv-year
Metrics: node classification accuracy
Baselines: FedAvg FedProx FedNova FedFa FedSage+ FedStar FedPub FGSSL FedGTA AdaFGL FGGP
Builds on: ACM-GCN FedAvg Fisher Information Matrix
Closest papers by meaning
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.92)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.90)
- Modeling Inter-Intra Heterogeneity for Graph Federated Learning (AAAI 2025 · similarity 0.90)
- FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.90)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.90)
- $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.