Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Graph Learning under Domain Shift with Generalizable Prototypes
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 FGGP, a federated graph learning framework for domain-shifted subgraphs that replaces classifiers with clustered prototype memory and uses global-knowledge-injected contrastive learning to improve multi-domain generalization.
Problem
Federated graph learning suffers degraded generalization of the global model when client graphs come from diverse domains with attribute and structure shifts.
Approach
Decouples the global model into a feature extractor and a classifier linked by prototypes: clients upload prototypes instead of classifier weights, the server clusters them with FINCH into a memory bank, and prediction uses kNN voting over cosine similarity. The feature extractor is trained with a global-knowledge-injected view (semantic kNN graph completion plus adaptive Gumbel-softmax topology)
Key results
- On Twitch with alpha=5, FGGP reaches 62.67% average accuracy, +12.26 over FedAvg (50.41%).
- On Citation with alpha=10, FGGP reaches 76.77% average accuracy on DBLP/ACM, +2.11 over FedAvg-level baseline (74.66%).
- Ablation on Twitch shows FCPP alone gives 59.80% and GKIC alone 50.04%, versus 61.74% with both.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FGGP instead of FedOPT | global federated GCN · Twitch α=5 (AVG) | accuracy: 58.94 -> 62.67 | Helped |
| use FGGP instead of FedOPT | global federated GCN · Citation α=5 (AVG) | accuracy: 72.35 -> 74.80 | Helped |
| use FGGP instead of FedOPT | global federated GCN · Twitch α=10 (AVG) | accuracy: 50.43 -> 61.74 | Helped |
| use FGGP instead of FedOPT | global federated GCN · Citation α=10 (AVG) | accuracy: 75.28 -> 76.77 | Helped |
| add FCPP (ablation) | FGGP with GKIC · Twitch α=10 (AVG) | accuracy: 50.04 -> 61.74 | Helped |
| add GKIC (ablation) | FGGP with FCPP · Twitch α=10 (AVG) | accuracy: 59.80 -> 61.74 | Helped |
| replace FINCH with Kmeans (ablation) | FGGP · Twitch α=5 | accuracy: 62.67 -> 62.37 | Hurt |
| replace FINCH with DBSCAN (ablation) | FGGP · Twitch α=5 | accuracy: 62.67 -> 61.33 | Hurt |
Surprising findings
- Several standard FL baselines (FedProx, MOON, FedProto) underperform FedAvg on some Twitch and Citation splits.
Limitations
- Evaluation limited to two graph datasets (Twitch, Citation) with GCN backbones.
- Client-side prototype memory and clustering depend on FINCH's parameter-free assumption, with no evaluation on unknown domain counts beyond these settings.
Experimental setup
Datasets: Twitch Citation (DBLP, ACM)
Benchmarks: Twitch Citation
Metrics: node classification accuracy average accuracy across domains
Baselines: FedAvg FedProx MOON FedOPT FedProc FedSage FedProto
Builds on: FedProto FedProc FINCH MOON SimGRACE
Closest papers by meaning
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.90)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.90)
- Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.90)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.90)
- FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.90)
- Subgraph Federated Learning via Spectral Methods (NeurIPS 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.