Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation
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 FedGOG, a federated graph OOD generalization framework combining score-based graph diffusion for OOD data exploration with mutual-information-based decorrelation between invariant and spurious GNN embeddings.
Problem
Federated graph learning degrades under out-of-distribution shifts between training and testing graphs at each client, with limited and heterogeneous client data making invariant learning hard.
Approach
Clients jointly train global score-based graph diffusion models to sample controllable OOD graphs beyond local data (DDE), and each client trains a global invariant GNN and a personalized spurious GNN whose latent embeddings are decorrelated via a mutual information estimate (LED).
Key results
- FedGOG reaches 40.74% on CMNIST (alpha=0.1) vs 20.40% for FedAvg and 35.11% for Fed-LECI, the strongest baseline there.
- FedGOG attains 59.65% on Motif (alpha=0.1) vs 53.09% for FedStar, and 73.69% on Motif (alpha=5.0) vs 71.46% for FedStar.
- Ablation on CMNIST (alpha=0.1): removing LED drops accuracy to 29.44%, removing DDE to 37.41%, full model 40.74%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · Motif, α=0.1 (class heterogeneity) | accuracy: 53.09±0.41 -> 59.65±1.04 | Helped |
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · CMNIST, α=0.1 (class heterogeneity) | accuracy: 35.11±1.51 -> 40.74±0.45 | Helped |
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · HIV, α=0.1 (class heterogeneity) | accuracy: 63.05±0.47 -> 64.26±0.80 | Helped |
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · LBAPcore, α=0.1 (class heterogeneity) | accuracy: 67.24±0.53 -> 70.66±0.39 | Helped |
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · Twitter, α=0.1 (class heterogeneity) | accuracy: 53.57±0.39 -> 57.98±0.53 | Helped |
| Use FedGOG instead of the strongest listed baseline | Federated graph classification · SST2, α=0.1 (class heterogeneity) | accuracy: 59.98±2.17 -> 64.77±1.72 | Helped |
| Add diffusion data exploration (DDE) (ablation) | FedGOG · HIV, α=0.1 (class heterogeneity) | accuracy: 63.98±0.67 -> 64.26±0.80 | Helped |
| Add diffusion data exploration (DDE) (ablation) | FedGOG · HIV, α=5.0 (class heterogeneity) | accuracy: 66.27±0.52 -> 67.38±0.68 | Helped |
| Add diffusion data exploration (DDE) (ablation) | FedGOG · CMNIST, α=0.1 (class heterogeneity) | accuracy: 37.41±1.03 -> 40.74±0.45 | Helped |
| Add diffusion data exploration (DDE) (ablation) | FedGOG · CMNIST, α=5.0 (class heterogeneity) | accuracy: 52.17±0.21 -> 55.88±0.11 | Helped |
| Add latent embedding decorrelation (LED) (ablation) | FedGOG · HIV, α=0.1 (class heterogeneity) | accuracy: 62.59±1.12 -> 64.26±0.80 | Helped |
| Add latent embedding decorrelation (LED) (ablation) | FedGOG · HIV, α=5.0 (class heterogeneity) | accuracy: 64.02±0.76 -> 67.38±0.68 | Helped |
Surprising findings
- FedIIR, a federated OOD method from image classification, does not significantly outperform FedAvg in the FGL OOD setting.
- Centralized graph OOD methods mostly improve only on the simple Motif dataset and can underperform FedAvg on other datasets, e.g., Fed-GIL on LBAPcore at alpha=0.1.
Limitations
- Performance drops notably as the number of clients increases (Fig. 5), suggesting scalability limits.
- Sampling-based OOD exploration requires training and aggregating diffusion score models, adding communication and computation overhead.
- Reported results use only five random seeds and standard GIN backbones; generality to other GNN architectures is not explored.
Experimental setup
Datasets: Motif CMNIST LBAPcore HIV Twitter SST2
Benchmarks: Motif CMNIST LBAPcore HIV Twitter SST2
Metrics: accuracy
Baselines: FedAvg FedIIR Fed-DIR Fed-GIL Fed-LECI Fed-AIA GCFL FedStar
Builds on: FedAvg score-based diffusion (SDE) MINE (Donsker-Varadhan mutual information estimator) GOOD benchmark FedStar LECI
Closest papers by meaning
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.90)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.90)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.89)
- Virtual Nodes Can Help: Tackling Distribution Shifts in Federated Graph Learning (AAAI 2025 · similarity 0.89)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.89)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 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.