Research map / Federated learning and privacy / Personalized and heterogeneous FL
Decoupled Subgraph 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 FedStruct, a subgraph federated learning framework that shares only global graph structure (via a decoupled GCN over learnable node structure features, Hop2Vec) instead of node features or embeddings, achieving near-centralized accuracy for semi-supervised node classification.
Problem
Subgraph federated learning for node classification where clients hold interconnected subgraphs and cannot share node features or embeddings without privacy risk, while maintaining accuracy and low communication cost.
Approach
Decouple node features from structure: each client computes node feature embeddings locally, while node structure embeddings are obtained from a decoupled GCN over learnable node structure features (Hop2Vec) using only the local partition of the L-hop combined adjacency matrix, which is precomputed privately with pruning; no node features or embeddings are exchanged.
Key results
- ...
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use task-dependent Hop2Vec NSF instead of one-hot degree NSF (ablation) | FedStruct · Cora, 10 clients, random partitioning | accuracy: 69.89±1.85 -> 79.27±0.90 | Helped |
| use task-dependent Hop2Vec NSF instead of FedStar NSF (ablation) | FedStruct · Chameleon, 10 clients, random partitioning | accuracy: 41.89±1.67 -> 52.60± 1.25 | Helped |
| disable pruning of the combined adjacency matrix (ablation) | FedStruct (H2V) · Cora, 10 clients, random partitioning | accuracy: 79.27±0.90 -> 79.88±0.92 | Helped |
| use FedStruct (H2V) instead of FedGCN-2hop | node classification · Chameleon, 10 clients, random partitioning | accuracy: 48.77± 1.73 -> 52.60± 1.25 | Helped |
| use FedStruct (H2V) instead of FedGCN-2hop | node classification · Amazon Ratings, 10 clients, random partitioning | accuracy: 40.78± 0.56 -> 40.97±0.64 | No clear effect |
| use FedStruct (H2V) instead of FedGCN-2hop | node classification · Cora, Citeseer, Pubmed, Amazon Photo; 10 clients, random partitioning | accuracy | Hurt |
| use FedStruct (H2V) instead of FedSage+ | node classification · Chameleon, 20 clients, random partitioning | accuracy: 34.33% -> 52.76% | Helped |
| use NSEs only rather than NSEs plus NFEs at inference (ablation) | FedStruct (H2V) · Cora, K-means partitioning, 5 clients | accuracy: 75% -> 80% | Hurt |
Limitations
- FedGCN baseline lacks privacy and is included only as an upper reference
- Hop2Vec pruned version slightly underperforms the non-pruned variant
- Filter parameters fixed to beta_Ls=1 without tuning
Experimental setup
Datasets: Cora Citeseer Pubmed Chameleon Amazon Photo Amazon Ratings
Benchmarks: Cora Citeseer Pubmed Chameleon Amazon Photo Amazon Ratings
Metrics: node classification accuracy
Baselines: FedSGD GNN FedSage+ FedPub FedGCN Local GNN Central GNN
Builds on: FedSage+ FedGCN FedStar GraphSAGE
Closest papers by meaning
- Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.92)
- Federated Node-Level Clustering Network with Cross-Subgraph Link Mending (ICML 2025 · similarity 0.90)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.90)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.90)
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.89)
- Federated Graph-Level Clustering Network (AAAI 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.