Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Graph-Level Clustering Network
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 FedGCN, an unsupervised federated graph-level clustering framework that shares clustering prototypes from clients and fuses them into consensus prototypes at the server via an improved GMM/KDE estimation.
Problem
Federated graph learning typically needs labeled graphs; in unsupervised, cross-domain non-IID graph settings, collaborative learning is hindered and sharing raw attributes risks privacy leakage.
Approach
Clients extract structure-oriented graph embeddings (FedStar-style degree and random-walk features with a GIN encoder), cluster them, and upload hard-to-reconstruct cluster prototypes; the server fuses them into consensus prototypes via a kernel-density-improved Gaussian mixture EM, which are sent back to clients as cluster centers for a KL-based clustering loss.
Key results
- On the SM (7-client, 2-class) setting FedGCN reaches 75.9% ACC, 24.9 NMI, 31.1 ARI, versus 58.9% ACC for FedStar (unsupervised adaptation).
- On the CV 15-cluster setting FedGCN achieves 34.6% ACC vs 13.7% for GCFL.
- Adding the prototype-sharing component improves FedAvg by an average of up to 7.2% (SM) and 3.9% (SM-BIO) across four metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add consensus prototype optimizing (CPO) to the baseline client model (ablation) | FedGCN client model · All datasets shown in Figure 3 | Clustering performance | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM (7) | ACC: 71.6±0.7 -> 75.9±0.8 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM (7) | NMI: 16.3±2.8 -> 24.9±3.0 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM (7) | ARI: 20.4±1.5 -> 31.1±3.4 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM (7) | F1: 62.1±1.0 -> 67.1±1.5 | Helped |
| Add prototype-sharing (PS) to FedProx (ablation) | FedProx · SM (7) | ACC: 73.3±0.8 -> 73.9±0.8 | Helped |
| Add prototype-sharing (PS) to FedPer (ablation) | FedPer · SM (7) | ARI: 21.7±1.6 -> 25.6±2.7 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM-BIO (9) | ACC: 66.7±0.2 -> 69.5±0.7 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM-BIO (9) | NMI: 11.1±1.3 -> 14.9±2.2 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM-BIO (9) | ARI: 13.8±1.1 -> 18.9±2.7 | Helped |
| Add prototype-sharing (PS) to FedAvg (ablation) | FedAvg · SM-BIO (9) | F1: 57.4±1.5 -> 61.3±1.3 | Helped |
| Add prototype-sharing (PS) to FedPer (ablation) | FedPer · SM-BIO (9) | NMI: 10.5±1.3 -> 11.7±1.9 | Helped |
Surprising findings
- Supervised FGL methods adapted to unsupervised clustering perform close to chance on several settings, e.g. FedSage ACC 10.1% on the 15-cluster CV setting.
- On the SM-BIO-SY setting FedGCN's NMI (13.5) is lower than FedSage's (20.6), so it is not best on all metrics.
Limitations
- Evaluated only on graph-level datasets with a fixed set of clients and non-IID partitions; no node-level or incomplete-graph settings.
- Supervised baselines were given a few labels and adapted to unsupervised settings, so comparisons are not like-for-like.
- Hyperparameters such as lambda and bandwidth require tuning; performance varies with lambda.
Experimental setup
Datasets: MUTAG BZR COX2 DHFR PTC_MR AIDS BZR_MD DD PROTEINS SYNTHETIC COLLAB IMDB-MULTI
Benchmarks: MUTAG BZR COX2 DHFR PTC_MR AIDS BZR_MD DD PROTEINS SYNTHETIC COLLAB IMDB-MULTI
Metrics: ACC NMI ARI F1
Baselines: FedAvg FedProx FedPer FedSage GCFL FedStar
Builds on: FedStar FedAvg GCFL GIN
Closest papers by meaning
- Federated Node-Level Clustering Network with Cross-Subgraph Link Mending (ICML 2025 · similarity 0.92)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.89)
- $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.89)
- Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.89)
- Federated Graph Condensation with Information Bottleneck Principles (AAAI 2025 · similarity 0.88)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 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.