Research map / Label-efficient and robust learning / Domain adaptation and generalization
DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start 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 DisCo, a graph-based disentangled intent-wise contrastive framework that uses affinity-graph random walks to filter irrelevant source-domain collaborative information for cold-start cross-domain recommendation.
Problem
Cold-start cross-domain recommendation suffers from negative transfer: users similar in the source domain may differ in the target domain, so transferring source embeddings injects irrelevant collaborative information.
Approach
A multi-channel graph encoder extracts K user intents per domain; multi-step random walks over an embedding-space affinity graph give high-order user similarities that supervise intra-domain and inter-domain intent-wise contrastive losses, with a cross-domain decoder preserving target-specific preferences.
Key results
- On Game-Video, DisCo reaches HR@10 9.54 vs 8.72 for the best prior baseline (CDRIB) in the Game domain; on Music-Movie, DisCo reaches HR@10 15.92 and 16.33 vs 14.72 and 14.86 for CDRIB.
- On Sport-Cloth, DisCo is competitive but does not beat the best baseline: HR@10 10.72 vs 11.10 (CDRIB) on Sport and 12.85 vs 12.28 (CDRIB) on Cloth, with NDCG lower than CDRIB on Sport.
- Ablation: removing the cross-domain decoder, uniform intent prior, orthogonality loss, or affinity-graph random walks each reduces performance on DisCo's full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DisCo instead of the strongest baseline | Cold-start cross-domain recommendation · Game-video and music-movie | HR@10 and NDCG@10: Game HR@10: 8.72±0.41 -> 9.54±0.28; Game NDCG@10: 4.63±0.17 -> 4.87±0.27; Video HR@10: 12.66±0.39 -> 13.38±0.31; Video N | Helped |
| Use DisCo instead of the strongest baseline | Cold-start cross-domain recommendation · Sport-cloth and phone-electronic | HR@10 and NDCG@10: Sport HR@10: 11.10±0.29 -> 10.72±0.32; Sport NDCG@10: 6.19±0.17 -> 5.81±0.26; Cloth HR@10: 12.28±0.24 -> 12.85±0.42; Clo | Mixed |
| Add the cross-domain decoder (ablation) | DisCo · Sport, Cloth, Game, Video, Music, Movie, Phone, Elec | HR@10: Sport: 9.81±0.24 -> 10.72±0.32; Cloth: 11.79±0.29 -> 12.85±0.42; Game: 9.20±0.32 -> 9.54±0.28; Video: 11.5±0.31 -> 13.38 | Helped |
| Use user-specific intent weights instead of a uniform intent distribution (ablation) | DisCo · Sport, Cloth, Game, Video, Music, Movie, Phone, Elec | HR@10: Sport: 10.64±0.19 -> 10.72±0.32; Cloth: 12.55±0.20 -> 12.85±0.42; Game: 9.29±0.35 -> 9.54±0.28; Video: 12.06±0.33 -> 13. | Helped |
| Add orthogonality regularization across intents (ablation) | DisCo · Sport, Cloth, Game, Video, Music, Movie, Phone, Elec | HR@10: Sport: 10.62±0.37 -> 10.72±0.32; Cloth: 12.2±0.26 -> 12.85±0.42; Game: 9.41±0.24 -> 9.54±0.28; Video: 12.26±0.29 -> 13.3 | Helped |
| Use affinity graph construction and multi-step random walks rather than setting the walk transition to identity (ablation) | DisCo · Sport, Cloth, Game, Video, Music, Movie, Phone, Elec | HR@10: Sport: 9.56±0.27 -> 10.72±0.32; Cloth: 11.88±0.31 -> 12.85±0.42; Game: 8.62±0.26 -> 9.54±0.28; Video: 12.4±0.23 -> 13.38 | Helped |
Surprising findings
- On Sport-Cloth and Phone-Electronic, DisCo is only competitive with, and in some metrics trails, CDRIB and UniCDR, rather than consistently outperforming them despite the abstract's claim.
Limitations
- Gains over strongest baselines are inconsistent across the four domain pairs, with Sport-Cloth and Phone-Electronic showing only competitive results.
- Optimal number of intents and random-walk steps vary by domain and must be tuned by grid search.
- Evaluation uses only one source-target pairing setup with 20% of overlapping users as cold-start users.
Experimental setup
Datasets: Amazon music-movie Amazon phone-electronic Amazon cloth-sport Amazon game-video
Benchmarks: Amazon music-movie Amazon phone-electronic Amazon cloth-sport Amazon game-video
Metrics: HR@10 NDCG@10
Baselines: CML BPR-MF NGCF EMCDR SSCDR TMCDR SA-VAE CDRIB UniCDR
Builds on: EMCDR CDRIB UniCDR DGCL NGCF
Closest papers by meaning
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.94)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.91)
- Disentangled Contrastive Bundle Recommendation with Conditional Diffusion (AAAI 2025 · similarity 0.88)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.87)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.87)
- Federated Graph Learning for Cross-Domain Recommendation (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.