Research map / Label-efficient and robust learning / Domain adaptation and generalization
Federated Graph Learning for Cross-Domain Recommendation
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 FedGCDR, a federated graph-attention framework for broader-source cross-domain recommendation that uses Gaussian-perturbed knowledge transfer with MLP feature mapping and virtual-social-link graph expansion to filter negative transfer, evaluated on 4/8/16 Amazon domains.
Problem
Cross-domain recommendation with many source domains faces privacy leakage (intra- and inter-domain) and accumulating negative transfer as the number of source domains grows.
Approach
A horizontal-vertical-horizontal federated pipeline in which source GAT embeddings are perturbed with Gaussian noise and mapped by per-domain MLPs, then the target graph is expanded with virtual users built from source embeddings so GAT attention yields domain attentions; a final fine-tuning stage freezes propagation layers to suppress noise-induced negative transfer.
Key results
- On Amazon-16@Books, FedGCDR reaches HR@10 0.6516 and NDCG@10 0.3854 versus best baseline FedCDR 0.5626 / 0.3403 (HR@10 / NDCG@10).
- On Amazon-16@CDs, FedGCDR achieves HR@5 0.5267 and NDCG@5 0.4382, outperforming FedCDR's 0.4823 / 0.3983.
- On dual-domain Books→CDs, FedGCDR reaches HR@10 0.3323 versus 0.2903 for PriCDR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedGCDR instead of the strongest baseline | Amazon-4@Books · Amazon-4@Books | HR@5: 0.4693 -> 0.4941 | Helped |
| Use FedGCDR instead of the strongest baseline | Amazon-8@Books · Amazon-8@Books | HR@5: 0.4791 -> 0.5300 | Helped |
| Use FedGCDR instead of the strongest baseline | Amazon-16@Books · Amazon-16@Books | HR@5: 0.4693 -> 0.5016 | Helped |
| Use FedGCDR instead of the strongest baseline | Amazon-4@CDs · Amazon-4@CDs | HR@5: 0.4299 -> 0.4588 | Helped |
| Use FedGCDR instead of the strongest baseline | Amazon-8@CDs · Amazon-8@CDs | HR@5: 0.4119 -> 0.4276 | Helped |
| Use FedGCDR instead of the strongest baseline | Amazon-16@CDs · Amazon-16@CDs | HR@5: 0.4823 -> 0.5267 | Helped |
| Add Gaussian noise to source embeddings (ablation) | FedGCDR · Amazon-16@CDs | HR@5: 0.5267 -> 0.4963 | Hurt |
| Add Gaussian noise to source embeddings (ablation) | FedGCDR · Amazon-16@Books | HR@5: 0.5016 -> 0.4928 | Hurt |
| Replace attention graph expansion with averaged source embeddings and omit fine-tuning (ablation) | FedGCDR · Amazon-16@Books | recommendation performance | Hurt |
| Transfer source embeddings without feature mapping (ablation) | FedGCDR · Amazon-16@CDs | recommendation performance | Hurt |
| Add feature mapping while retaining attention graph expansion and fine-tuning (ablation) | FedGCDR · Amazon-16@Books | recommendation performance | Mixed |
| Use the positive knowledge activation module rather than averaged embeddings and no fine-tuning (ablation) | FedGCDR · Amazon-16@Books | recommendation performance | Helped |
Surprising findings
- Several CDR baselines (EMCDR, PriCDR, FedCT, FedCDR) perform worse with 16 source domains than with 4 or 8, indicating hard negative transfer as domains are added.
- On the Books target, ablating the transfer-mapping module (FedGCDR-T) hurt less than ablating the activation module, while for CDs the opposite held.
Limitations
- Evaluated only on the Amazon dataset, limiting generalizability to other data sources.
- Relies on overlapping users as the cross-domain bridge.
- Privacy-utility trade-off: smaller epsilon degrades performance noticeably.
Experimental setup
Datasets: Amazon 2018 (16 domains) Amazon-Dual
Benchmarks: Amazon-4 Amazon-8 Amazon-16 Amazon-Dual
Metrics: HR@5 NDCG@5 HR@10 NDCG@10
Baselines: FedGNN (Single Domain) EMCDR PriCDR FedCT FedCDR
Builds on: LightGCN EMCDR FedGNN FedCDR PriCDR
Closest papers by meaning
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.87)
- Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation (ICML 2024 · similarity 0.87)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.87)
- Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.87)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.86)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · 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.