Research map / Graph learning / Graph self-supervised learning
Modeling Inter-Intra Heterogeneity for Graph Federated 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 FedIIH, a graph federated learning method that infers subgraph data distributions with a hierarchical variational graph autoencoder for inter-subgraph similarity and disentangles each subgraph into latent factors with per-factor separate federation for intra-heterogeneity.
Problem
Personalized graph federated learning suffers from unreliable inter-subgraph similarity estimated from local model outputs and ignores heterogeneity within each subgraph (intra-heterogeneity).
Approach
A hierarchical variational graph autoencoder (HVGAE) links local latent factors to shared global latent factors so that JS divergence between inferred subgraph distributions gives per-factor client similarities; a DisenGCN disentangles each subgraph into K factors and parameters are federated separately per factor.
Key results
- On heterophilic datasets FedIIH outperforms the second-best method (FedSage+) by 5.79% in the non-overlapping setting and 4.53% in the overlapping setting.
- On homophilic datasets FedIIH reaches 83.10% average accuracy in the non-overlapping setting, 1.51% above FED-PUB.
- Ablation on Cora (10 clients, non-overlapping) shows removing variational inference drops accuracy from 81.85% to 73.51%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedIIH instead of FED-PUB | Graph federated node classification on homophilic datasets · Non-overlapping, all datasets | average accuracy: 83.10 vs 81.59; 1.51% higher | Helped |
| Use FedIIH instead of FED-PUB | Graph federated node classification on homophilic datasets · Overlapping, all datasets | average accuracy: 81.01 vs 79.53; 1.48% higher | Helped |
| Use FedIIH instead of FedSage+ | Graph federated node classification on heterophilic datasets · Non-overlapping, all datasets | average score: 64.52 vs 58.73; 5.79% | Helped |
| Use FedIIH instead of FedSage+ | Graph federated node classification on heterophilic datasets · Overlapping, all datasets | average score: 62.50 vs 57.97; 4.53% | Helped |
| Add the hierarchical model (ablation) | FedIIH · Cora and Amazon-ratings, non-overlapping and overlapping | accuracy: Cora: ↓3.18, ↓1.99; Amazon-ratings: ↓2.55, ↓4.59 when removed | Helped |
| Add variational inference (ablation) | FedIIH · Cora and Amazon-ratings, non-overlapping and overlapping | accuracy: Cora: ↓8.34, ↓2.17; Amazon-ratings: ↓2.56, ↓1.93 when removed | Helped |
| Add disentanglement (ablation) | FedIIH · Cora and Amazon-ratings, non-overlapping and overlapping | accuracy: Cora: ↓3.07, ↓3.39; Amazon-ratings: ↓3.04, ↓2.65 when removed | Helped |
| Increase the number of latent factors K (ablation) | FedIIH · Tolokers | performance | Helped |
| Use K=1 instead of multiple latent factors (ablation) | FedIIH · Most datasets | performance | Hurt |
Surprising findings
- Removing the variational inference component causes a large drop on Cora non-overlapping (8.34 points) but a small drop in the overlapping setting (2.17 points).
Limitations
- Intra-heterogeneity is modeled only with DisenGCN, chosen for simplicity; other disentangled GCNs are not explored.
- Gains on homophilic datasets are modest (about 1.5 points).
Experimental setup
Datasets: Cora CiteSeer PubMed Amazon-Computer Amazon-Photo ogbn-arxiv Roman-empire Amazon-ratings Minesweeper Tolokers Questions
Benchmarks: Cora CiteSeer PubMed Amazon-Computer Amazon-Photo ogbn-arxiv Roman-empire Amazon-ratings Minesweeper Tolokers Questions
Metrics: accuracy AUC
Baselines: Local FedAvg FedProx FedPer GCFL FedGNN FedSage+ FED-PUB FedGTA AdaFGL
Builds on: FED-PUB DisenGCN VGAE FedHB
Closest papers by meaning
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.90)
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.90)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.89)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.89)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.88)
- Subgraph Federated Learning for Local Generalization (ICLR 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.