Research map / Federated learning and privacy / Personalized and heterogeneous FL
Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement
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 FedGFM, a federated paradigm for training graph foundation models across silos, and FedGFM+, which mitigates "knowledge entanglement" via anchor-based domain-aware codebook initialization (AncDAI) and an adaptive domain-sensitive prompt pool (AdaDPP).
Problem
Training graph foundation models (GFMs) in a federated, cross-domain, cross-task setting, where naive federated pre-training causes domain-specific knowledge to collapse into indistinguishable representations (knowledge entanglement).
Approach
Initialize the global gVQ-VAE codebook from Gaussian-perturbed domain prototypes (mean-pooled client embeddings used as semantic anchors), and keep per-client graph prompts local during pre-training then pool them at fine-tuning with attention-based selection for the target graph.
Key results
- FedGFM+ reaches 83.79 on Cora, 88.52 on PubMed, and 76.31 on OGB-arxiv, outperforming the best listed baselines on all 8 datasets.
- Ablation: removing AncDAI drops Cora accuracy from 83.79 to 81.55 and HIV from 69.39 to 67.52; removing AdaDPP causes smaller drops (e.g., Cora 83.17).
- Federated naive GFT* shows near-unity inter-domain embedding similarity, indicating knowledge entanglement; FedGFM+ improves over 20 baselines including FL/FGL methods and federated GFM variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · Cora | accuracy: 82.41 ± 0.33 -> 83.79 ± 0.27 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · PubMed | accuracy: 87.10 ± 0.25 -> 88.52 ± 0.31 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · OGB-arxiv | accuracy: 74.19 ± 0.05 -> 76.31 ± 0.18 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · WikiCS | accuracy: 78.81 ± 0.19 -> 80.70 ± 0.28 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · FB15K237 | accuracy: 73.80 ± 0.48 -> 75.25 ± 0.24 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · WN18RR | accuracy: 86.44 ± 0.29 -> 89.25 ± 0.13 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · HIV | AUC: 67.93 ± 0.24 -> 69.39 ± 0.44 | Helped |
| use FedGFM+ instead of the strongest listed baseline | graph foundation model fine-tuning · PCBA | AUC: 73.51 ± 0.24 -> 77.68 ± 0.22 | Helped |
| add AncDAI to FedGFM+ (ablation) | FedGFM+ · Cora | accuracy: 81.55 ± 0.22 -> 83.79 ± 0.27 | Helped |
| add AncDAI to FedGFM+ (ablation) | FedGFM+ · PubMed | accuracy: 85.56 ± 0.28 -> 88.52 ± 0.31 | Helped |
| add AncDAI to FedGFM+ (ablation) | FedGFM+ · OGB-arxiv | accuracy: 75.19 ± 0.19 -> 76.31 ± 0.18 | Helped |
| add AdaDPP to FedGFM+ (ablation) | FedGFM+ · Cora | accuracy: 83.17 ± 0.18 -> 83.79 ± 0.27 | Helped |
Surprising findings
- Naive federated variants of centralized GFMs often fall below isolated supervised baselines (negative transfer).
- Under FedGFM+ the AdaDPP prompt count is insensitive, performing well with few prompts.
Limitations
- Privacy leakage risk from sharing prototypes and prompts is not formally analyzed.
- Experiments use a fixed partition into 3 clients per dataset with simulated decentralization.
- Theoretical results rely on informal assumptions (e.g., minimal angular separation, small sigma) and the proofs are loosely argued.
Experimental setup
Datasets: Cora PubMed OGB-arxiv WikiCS FB15K237 WN18RR HIV PCBA
Benchmarks: Cora PubMed OGB-arxiv WikiCS FB15K237 WN18RR HIV PCBA
Metrics: accuracy AUC
Baselines: Linear GCN GAT GraphSAGE GIN FedAvg MOON FedSage+ Fed-PUB FedGTA FedTAD FGSSL
Builds on: GFT FedAvg MOON Sentence-BERT Louvain
Closest papers by meaning
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.90)
- FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.90)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.89)
- Modeling Inter-Intra Heterogeneity for Graph Federated Learning (AAAI 2025 · similarity 0.89)
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.89)
- Subgraph Federated Learning for Local Generalization (ICLR 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.