Research map / Graph learning / Graph self-supervised learning
Virtual Nodes Can Help: Tackling Distribution Shifts in Federated Graph Learning
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 FedVN, a federated graph learning framework that adds multiple shared learnable virtual nodes whose edges to local graphs are set by a personalized edge generator, so that graphs become client-invariant under distribution shift.
Problem
Graph property prediction in federated graph learning suffers from distribution shifts across clients, where label-irrelevant substructures differ per client and degrade GNN training.
Approach
Clients jointly learn a set of shared virtual nodes (VNs) while each client trains a personalized edge generator that decides how VNs connect its local graphs; a decoupling loss keeps VNs diverse and a score-contrastive loss makes edge patterns consistent within a client but distinct across clients.
Key results
- FedVN reaches 75.72% accuracy on Motif/Basis versus 67.12% for the best baseline (Self-training) and 63.62% for FedStar.
- On ZINC/Scaffold FedVN achieves 0.4947 MAE, lower than the best baseline FedBN at 0.5129.
- Removing the edge generator drops FedVN from 75.72% to 60.20% on Motif/Basis and from 43.67% to 40.15% on CMNIST/Color.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedVN instead of Self-training | GIN for federated graph learning · Motif/Basis | accuracy: 67.12$\pm$0.89 -> 75.72$\pm$1.85 | Helped |
| use FedVN instead of GCFL+ | GIN for federated graph learning · Motif/Size | accuracy: 49.34$\pm$2.70 -> 50.41$\pm$1.17 | Helped |
| use FedVN instead of FedProx | GIN for federated graph learning · CMNIST/Color | accuracy: 39.78$\pm$0.68 -> 43.67$\pm$1.25 | Helped |
| use FedVN instead of FedBN | GIN for federated graph learning · ZINC/Scaffold | MAE: 0.5129$\pm$0.0119 -> 0.4947$\pm$0.0174 | Helped |
| use FedVN instead of FedAvg | GIN for federated graph learning · SST2/Length | accuracy: 81.79$\pm$0.27 -> 83.13$\pm$0.79 | Helped |
| add personalized edge generators (ablation) | FedVN · Motif/Basis | accuracy: 60.20$\pm$1.75 -> 75.72$\pm$1.85 | Helped |
| add personalized edge generators (ablation) | FedVN · CMNIST/Color | accuracy: 40.15$\pm$1.27 -> 43.67$\pm$1.25 | Helped |
| set the decoupling-loss weight to 0 (ablation) | FedVN · SST2/Length | accuracy | Hurt |
| set the score-contrastive-loss weight to 0 (ablation) | FedVN · SST2/Length | accuracy | Hurt |
| remove the decoupling loss (ablation) | FedVN · ZINC/Scaffold and SST2/Length | pairwise VN cosine similarity | Hurt |
| vary the number of virtual nodes; use only 1 or 5 (ablation) | FedVN · Motif/Size and CMNIST/Color | performance | Hurt |
Surprising findings
- FedStar degrades sharply on CMNIST/Color (28.10%) and SST2/Length (58.57%), which the authors attribute to sharing structure encoders when labels depend on node features.
Limitations
- Adds extra computation and communication cost from edge generators and multiple VNs.
- Shared VNs may leak private information and are vulnerable to malicious clients perturbing them; privacy not analyzed.
- Theoretical result (Theorem 1) is shown for a linear/single-layer GIN setting and extended to multi-layer by informal argument.
Experimental setup
Datasets: Motif CMNIST ZINC SST2
Benchmarks: Motif/Basis Motif/Size CMNIST/Color ZINC/Scaffold SST2/Length
Metrics: accuracy MAE
Baselines: Self-training FedAvg FedProx FedBN Ditto FedRep FedALA GCFL+ FedStar
Builds on: FedAvg GIN NT-Xent FedBN
Closest papers by meaning
- $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.92)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.90)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.89)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.88)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.88)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.88)
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.