Research map / Graph learning / Graph self-supervised learning

Federated Node-Level Clustering Network with Cross-Subgraph Link Mending

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

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 FedNCN, an unsupervised federated node-level clustering framework that mends cross-subgraph links on the server using clustering prototypes uploaded by clients via an MLP projector, N-Cut pruning, and a GNN generator.

Problem

Node-level federated graph learning relies on labels and, when a complete graph is partitioned into subgraphs across clients, loses cross-subgraph links, degrading node representations for clustering.

Approach

Clients upload hard-to-reconstruct clustering signals (Gaussian-noise masked matrix plus a trained MLP projector) instead of raw node data; the server estimates cross-subgraph affinities via graph kernels, prunes redundant edges with an improved N-Cut criterion, and trains a GNN generator on the mended graph to produce consensus prototypes shared back to clients.

Key results

• On CiteSeer with 5/10/20 clients, FedNCN reaches 54.98/58.99/57.47 ACC, versus 24.05/26.34/28.05 for the best baseline FedGTA.
• Cross-subgraph link mending raises the similarity of restored topology from 44.74% to 78.45% on CiteSeer (Fig. 1).
• Average metric gains over local training at 10 clients reach +18.0 on Amazon-Computer and +22.4 on Amazon-Photo, versus at most +3.2 for classical FL aggregation baselines.

Measured effects reported in the paper

Surprising findings

• FedAvg yields lower NMI and ARI than the local-only model on CiteSeer and PubMed, showing that naive aggregation can hurt clustering.

Limitations

• Evaluated only on five benchmark datasets and simulated partitions produced by splitting a graph into 5, 10, or 20 clients.
• Performance is sensitive to the sampling ratio k, with both too few and too many uploaded samples degrading results.
• Baselines are supervised FGL methods adapted to the unsupervised setting, which may weaken the comparison.

Experimental setup

Datasets: CiteSeer PubMed Amazon-Computer Amazon-Photo Questions

Benchmarks: CiteSeer PubMed Amazon-Computer Amazon-Photo Questions

Metrics: ACC NMI ARI F1

Baselines: FedSage+ FedPUB FedTAD FedGTA FedIIH FedAvg FedProx FedPer

Builds on: FedTAD FedPUB DFCN K-means

Closest papers by meaning

• Federated Graph-Level Clustering Network (AAAI 2025 · similarity 0.92)
• Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.90)
• Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.88)
• Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections (NeurIPS 2025 · similarity 0.87)
• Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.87)
• EAGLES: Towards Effective, Efficient, and Economical Federated Graph Learning via Unified Sparsification (ICML 2025 · similarity 0.86)