Research map / Federated learning and privacy / Personalized and heterogeneous FL
Large Language Models Enhanced Personalized Graph Neural Architecture Search in 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 PFGNAS, which uses an LLM controller guided by task-specific prompts (with graph statistics and performance feedback) to generate personalized GNN architectures for federated clients, with a weight-sharing supernet pruned per client to counter client drift.
Problem
Automating personalized GNN architecture design in federated learning on graphs, where heterogeneous client data cause client drift and expert design effort is high.
Approach
Use an LLM as a self-reflective architecture controller with graph-statistics-aware prompts and historical performance feedback to propose per-client GNN architectures, while a shared supernet with per-client pruning masks and FedAvg-style weight sharing handles personalization and heterogeneity.
Key results
- On Cora and Citeseer under uniform partitioning, PFGNAS reports 83.80% (Cora, N=3) and 80.70% (Citeseer, N=3) accuracy, outperforming FedPUB on all metrics per the authors.
- On PubMed with heterogeneous partitioning (N=3, beta=0.2), PFGNAS's final global accuracy reaches 86.13%, a 19.42% improvement over PFGNAS-O and 6.71% over PFGNAS-E.
- Table 2 reports a 23.75% average improvement over the FL+GCN baseline on PubMed (N=3, beta=0.2), versus 12.88% for FedPUB and -33.74% for FL+GraphNAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PFGNAS instead of FedPUB | personalized federated GNN architecture search · Cora, N=3 | accuracy: 82.32±0.62 -> 83.80±2.22 | Helped |
| use PFGNAS instead of FLAGNNS | personalized federated GNN architecture search · Cora, N=5 | accuracy: 82.20±1.27 -> 80.86±1.73 | Hurt |
| use PFGNAS instead of SGC | personalized federated GNN architecture search · Citeseer, N=3 | accuracy: 80.20±0.40 -> 80.70±0.95 | Helped |
| use PFGNAS instead of GAT | personalized federated GNN architecture search · Citeseer, N=5 | accuracy: 78.63±1.24 -> 79.79±1.63 | Helped |
| replace personalized client architectures with one uniform client model (ablation) | PFGNAS · PubMed, N=3, β=0.2, Global | accuracy: 0.73 -> 0.86 | Hurt |
| replace LLM-guided architecture updates with top-5 historical architecture crossover (ablation) | PFGNAS · PubMed, N=3, β=0.2, Global | accuracy: 0.81 -> 0.86 | Hurt |
| use LLM-guided architecture recommendations instead of random client architecture assignment (ablation) | PFGNAS · PubMed, heterogeneous data distributions | accuracy | Helped |
Surprising findings
- FL+GraphNAS performed worse than the FL+GCN baseline on PubMed under heterogeneous partitioning, with Client 1 test accuracy at 0.00.
Limitations
- The method does not consistently achieve the best results across all metrics, as acknowledged by the authors.
- Evaluation limited to three citation graph datasets with node classification and a fixed search space of nine GNN operators.
- Results are averaged over three repetitions with considerable variance on some settings.
Experimental setup
Datasets: Cora Citeseer PubMed
Benchmarks: Cora Citeseer PubMed
Metrics: accuracy AUC
Baselines: GCN GraphSage GAT GIN SGC ARMA APPNP FL+DARTS FL+GraphNAS FLAGNNS FedPUB
Builds on: FedAvg DARTS GraphNAS GENIUS GPT4GNAS
Closest papers by meaning
- PerFedRLNAS: One-for-All Personalized Federated Neural Architecture Search (AAAI 2024 · similarity 0.90)
- FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.89)
- AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization (ICML 2025 · similarity 0.88)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.88)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.87)
- pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.86)
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.