Research map / Graph learning / Graph self-supervised learning
FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference
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 FedSSP, a personalized federated graph learning framework for graph classification under cross-domain structural shift, sharing spectral encoders (GSKS) and adding learnable client-specific preference adjustments with a feature-mean alignment regularizer (PGPA).
Problem
Personalized federated graph learning suffers from structural heterogeneity across domains: sharing non-generic structural knowledge causes knowledge conflict, and a single aggregated GNN fails to match each client's preferred message-passing scheme.
Approach
Share only spectral encoders (eigenvalue and filter encoders) that carry domain-generic knowledge while keeping the attention/decoder local, and add a learnable per-client preference vector to graph features regularized by MSE toward the global mean graph feature.
Key results
- FedSSP reaches 79.62% accuracy on the single-domain SM setting, versus 78.63% for the strongest baseline FedStar.
- On the SM-CV double-domain setting FedSSP reaches 84.29%, versus 82.60% for FGSSL.
- Ablation on SM shows GSKS alone gives 77.83% and PGPA alone 74.59%, versus 74.12% for the FedAvg-style baseline and 79.62% for the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedSSP instead of the strongest baseline | Federated graph classification · SM, SM-BIO, SM-CV, BIO-SN-CV, SM-SN-CV; compared with SM-BIO-SN | accuracy: 79.62 ± 2.23 vs 78.63 ± 2.11; 73.66 ± 2.34 vs 72.71 ± 1.22; 84.29 ± 0.68 vs 82.60 ± 0.48; 72.37 ± 2.18 vs 72.60 ± 2.45; | Mixed |
| Add GSKS to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM | accuracy: 74.12 -> 77.83 | Helped |
| Add GSKS to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM-CV | accuracy: 81.21 -> 82.78 | Helped |
| Add GSKS to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM-SN-CV | accuracy: 75.33 -> 78.54 | Helped |
| Add PGPA to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM | accuracy: 74.12 -> 74.59 | No clear effect |
| Add PGPA to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM-CV | accuracy: 81.21 -> 81.33 | No clear effect |
| Add PGPA to the FedSSP model without GSKS or PGPA (ablation) | FedSSP · SM-SN-CV | accuracy: 75.33 -> 76.12 | Helped |
| Add PGPA to GSKS (ablation) | FedSSP · SM | accuracy: 77.83 -> 79.62 | Helped |
| Add PGPA to GSKS (ablation) | FedSSP · SM-CV | accuracy: 82.78 -> 84.29 | Helped |
| Add PGPA to GSKS (ablation) | FedSSP · SM-SN-CV | accuracy: 78.54 -> 79.12 | Helped |
| Share only non-generic spectral components instead of using GSKS (ablation) | GSKS · SM | accuracy: 77.83 -> 77.21 | Hurt |
| Share only non-generic spectral components instead of using GSKS (ablation) | GSKS · SM-CV | accuracy: 82.78 -> 81.64 | Hurt |
Surprising findings
- Traditional FL methods FedAvg and FedProx failed to outperform local self-training in these cross-domain settings.
- Sharing all components or only non-generic spectral components often fails to beat self-training, while sharing only generic spectral encoders performs best.
Limitations
- Spectral GNNs require eigenvectors and eigenvalues, adding preprocessing and storage overhead.
- Evaluated only on graph classification with 15 small public datasets.
Experimental setup
Datasets: MUTAG BZR COX2 DHFR PTC_MR AIDS NCI1 PROTEINS OHSU Peking_1 IMDB-BINARY IMDB-MULTI
Benchmarks: SM SM-BIO SM-CV SM-BIO-SN BIO-SN-CV SM-SN-CV
Metrics: average test accuracy
Baselines: Local FedAvg FedProx FedSage GCFL APPLE FedCP FGSSL FedStar
Builds on: FedAvg FedProx APPLE FedCP FedStar GCFL
Closest papers by meaning
- $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.91)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.90)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.90)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.90)
- Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.90)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.90)
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.