Research map / Federated learning and privacy / Personalized and heterogeneous FL
Subgraph Federated Learning for Local Generalization
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 FedLoG, a subgraph federated learning method that condenses reliable head-degree, head-class knowledge into global synthetic node features and uses them to counter local overfitting for unseen nodes, missing classes, and new clients.
Problem
Subgraph federated learning on graphs suffers from local overfitting when new nodes, shifted label distributions, or missing classes appear at clients, so local models fail to generalize to unseen data.
Approach
Condense reliable knowledge (head-degree nodes at clients, weighted by head-class proportions at the server) into a small set of global synthetic node features, then train local models on these synthetic data after local fitting, with feature scaling and prompt-node generation to bridge the gap between synthetic and real graph nodes.
Key results
- In the Missing Class setting on Cora with 3 clients, FedLoG reaches 0.6472 accuracy versus 0.3900 for FedAvg and 0.3714 for FedNTD.
- On Seen Graph settings FedLoG attains the best accuracy in most configurations, e.g., 0.8601 on Cora with 3 clients versus 0.8622 for FedPUB at 10 clients where FedPUB is best.
- FedLoG reaches 0.8 accuracy on Cora with 3 clients in 10 rounds, versus 58 for FedAvg, though its per-round communication cost is higher (10.08 MB per round).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedLoG's reliable global synthetic data for local generalization | Subgraph federated learning · Unseen Node | accuracy | Helped |
| Use FedLoG's reliable global synthetic data for local generalization | Subgraph federated learning · Missing Class | accuracy | Helped |
| Use FedLoG | Subgraph federated learning · Seen Graph | accuracy | Helped |
Surprising findings
- Local training and FedAvg score near zero on the Missing Class setting for several datasets, showing that missing classes are nearly unlearnable without knowledge sharing.
- FedPUB and FedNTD sometimes outperform FedLoG on Seen Graph settings, e.g., FedPUB on Cora with 10 clients.
Limitations
- Higher per-round communication cost than most baselines.
- Evaluated only on five node-classification datasets with METIS-partitioned subgraphs, with a small number of clients.
Experimental setup
Datasets: Cora CiteSeer PubMed Amazon Computers Amazon Photo
Benchmarks: Cora CiteSeer PubMed Amazon Photo Amazon Computers
Metrics: node classification accuracy
Baselines: Local FedAvg FedSAGE+ FedGCN FedPUB FedNTD FedED
Builds on: GraphSAGE prototypical networks FedAvg METIS
Closest papers by meaning
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.90)
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.90)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.90)
- Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.90)
- $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.89)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.89)
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.