Research map / Label-efficient and robust learning / Domain adaptation and generalization
Graph Disentangled Contrastive Learning with Personalized Transfer 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 GDCCDR, a graph-disentangled contrastive framework for cross-domain recommendation that uses an adaptive parameter-free filter, item-aware proximate and eliminatory contrastive losses, and meta-network personalized transfer of domain-invariant user features.
Problem
Existing disentangled cross-domain recommenders treat all interactions equally, disentangle user features without regard to items, and transfer domain-invariant knowledge with non-personalized fusion, leading to sub-optimal disentanglement and transfer under data sparsity.
Approach
Weights each interaction in graph propagation via an adaptive parameter-free filter based on user-item similarity, uses two contrastive losses (proximate CL across domains for invariant features, eliminatory CL using user-item mutual information for specific features), and generates personalized transfer matrices with low-rank meta-networks.
Key results
- GDCCDR achieves the highest HR@10 and NDCG@10 across all four Amazon CDR datasets, e.g., 56.73 HR@10 / 37.96 NDCG@10 on Sport&Phone versus 52.57 / 35.87 for the runner-up BITGCF.
- Ablation removing eliminatory CL drops Sport&Phone HR@10 from 56.73 to 54.59 (w/o-ecl), and removing meta-networks drops Elec&Cloth HR@10 from 60.40 range to lower values.
- Reported p-values versus the runner-up are all below 0.01 across the four datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Sport&Phone (Sport) | HR@10: 52.57 -> 56.73 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Sport&Phone (Phone) | HR@10: 58.70 -> 64.71 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Sport&Cloth (Sport) | HR@10: 54.58 -> 59.67 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Sport&Cloth (Cloth) | HR@10: 51.80 -> 54.73 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Elec&Phone (Elec) | HR@10: 62.98 -> 64.58 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Elec&Phone (Phone) | HR@10: 65.03 -> 68.73 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Elec&Cloth (Elec) | HR@10: 58.78 -> 60.40 | Helped |
| Use GDCCDR instead of the strongest baseline | Cross-domain recommendation · Elec&Cloth (Cloth) | HR@10: 46.03 -> 49.83 | Helped |
| Add eliminatory contrastive learning (ablation) | GDCCDR · Sport&Phone (Sport) | HR@10: 54.59 -> 56.73 | Helped |
| Add proximate contrastive learning (ablation) | GDCCDR · Sport&Phone (Sport) | HR@10: 55.89 -> 56.73 | Helped |
| Add the adaptive parameter-free filter (ablation) | GDCCDR · Sport&Phone (Sport) | HR@10: 55.91 -> 56.73 | Helped |
| Use personalized meta-network transfer instead of average pooling (ablation) | GDCCDR · Sport&Phone (Sport) | HR@10: 55.47 -> 56.73 | Helped |
Surprising findings
- Standard CL-based CDR methods that align all user features across domains may introduce flaws by erasing domain-specific information, motivating the paper's separate treatment of invariant and specific features.
Limitations
- Evaluated only on four Amazon datasets with a single domain pair family (Sport/Elec with Phone/Cloth).
- Many hyperparameters (lambda_p, lambda_e, L, alpha, beta) require tuning, as shown in sensitivity analysis.
- Similarity function simplified to dot product; alternative similarity measures not explored.
Experimental setup
Datasets: Amazon Sport&Phone Amazon Sport&Cloth Amazon Elec&Phone Amazon Elec&Cloth
Benchmarks: Amazon Sport&Phone Amazon Sport&Cloth Amazon Elec&Phone Amazon Elec&Cloth
Metrics: HR@10 NDCG@10
Baselines: BPR NCF LightGCN DCCF CoNet DDTCDR DML BITGCF DisenCDR MADD ETL DCCDR
Builds on: DCCDR DisenCDR BITGCF LightGCN InfoNCE
Closest papers by meaning
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.94)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.89)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.88)
- Disentangled Contrastive Bundle Recommendation with Conditional Diffusion (AAAI 2025 · similarity 0.88)
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.87)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 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.