Research map / Graph learning / Graph self-supervised learning
EAGLES: Towards Effective, Efficient, and Economical Federated Graph Learning via Unified Sparsification
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 EAGLES, a unified graph-plus-parameter sparsification framework for federated graph learning that uses consensus-based parameter masks, a dual-expert graph sparsifier, and OT-based structure-similarity aggregation to cut training FLOPS and communication.
Problem
Federated graph learning incurs high computational and communication costs as graphs grow and GNNs deepen; existing sparsification methods address only graph or only parameter space and ignore client heterogeneity.
Approach
Jointly sparsify graph structure (via multiple sparsification experts combined by a hard-concrete-gated synergy expert) and model parameters (via client-consensus dynamic threshold masks with rollback), and weight client aggregation by optimal-transport distances between gating-derived structural representations.
Key results
- On Ogbn-Proteins with DeeperGCN, EAGLES reaches 82.10 ROC-AUC while reducing training FLOPS by 82% and communication bytes by 80% relative to baseline.
- On Pubmed with GCN, EAGLES reaches 86.97% top-1 accuracy, above FedAvg's 85.65%, at 0.48x FLOPS and 0.37x communication bytes.
- On Ogbn-Arxiv with GraphSAGE, EAGLES reaches 65.37% top-1 accuracy versus 63.99% for FedAvg, at 0.32x FLOPS and 0.48x communication bytes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GCN · Pubmed | Top-1 accuracy: 85.65 -> 86.97 | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GraphSAGE · Ogbn-Arxiv | Top-1 accuracy: 63.99 -> 65.37 | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | DeeperGCN · Ogbn-Proteins | ROC-AUC: 81.54 -> 82.10 | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GCN · Pubmed | Max training FLOPS: 1x -> 0.48x | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GraphSAGE · Ogbn-Arxiv | Max training FLOPS: 1x -> 0.32x | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | DeeperGCN · Ogbn-Proteins | Max training FLOPS: 1x -> 0.18x | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GCN · Pubmed | Communication bytes: 1x -> 0.37x | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | GraphSAGE · Ogbn-Arxiv | Communication bytes: 1x -> 0.48x | Helped |
| Use EAGLES unified graph and parameter sparsification instead of FedAvg | DeeperGCN · Ogbn-Proteins | Communication bytes: 1x -> 0.20x | Helped |
| Set parameter sparsification rate to 81.79% (ablation) | EAGLES · Ogbn-Proteins | Top-1 accuracy: +0.4% | Helped |
| Set graph sparsification rate to approximately 30% (ablation) | EAGLES · Multiple datasets | Top-1 accuracy | Mixed |
| Increase weight sparsity (ablation) | EAGLES · Ogbn-Arxiv and Ogbn-Proteins | Max training FLOPS | Helped |
Surprising findings
- Parameter sparsification at about 80% on Ogbn-Proteins slightly improves ROC-AUC (0.4%) over the unpruned baseline.
- ACE-GLT, a centralized dual-sparsification method, falls short of the baseline in several federated conditions.
Limitations
- Pretraining step inherited from lottery ticket methods adds upfront cost.
- Optimal graph sparsity reported around 30%, with parameter sparsity showing the main flexibility; sensitivity relies on a dynamic rollback heuristic.
- Gains over baselines are modest in several settings (e.g., about 1 point on Pubmed).
Experimental setup
Datasets: Cora Pubmed Photo Ogbn-Arxiv Ogbn-Proteins Ogbn-Products
Benchmarks: Cora Pubmed Photo Ogbn-Arxiv Ogbn-Proteins Ogbn-Products
Metrics: Top-1 accuracy ROC-AUC max training FLOPS communication bytes
Baselines: FedAvg FedProx FedSage+ APPLE FedCP FedTAD FedSSP FGGP FedGTA PruneFL FedTiny FedDIP
Builds on: lottery ticket hypothesis straight-through estimator hard concrete distribution DSpar ACE-GLT
Closest papers by meaning
- Graph Sparsification via Mixture of Graphs (ICLR 2025 · similarity 0.88)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.87)
- GRAIN: Exact Graph Reconstruction from Gradients (ICLR 2025 · similarity 0.87)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.87)
- Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.87)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.87)
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.