Research map / Label-efficient and robust learning / Recognition and long-tail
Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start 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 D2C2R, a domain-level disentanglement framework with variational graph autoencoders, attention-based information enhancement, and contrastive user personalized adaptation for zero-interaction cross-domain cold-start recommendation.
Problem
Providing recommendations for cold-start users who have no interactions in the target domain, using only their interactions from a source domain, in an end-to-end manner rather than via embedding mapping.
Approach
Independently extracts domain-specific knowledge with VGAE and shared knowledge with a fusion VGAE (FVGAE), fuses domains via an attention-based information enhancement network and matching-heuristic MLP mapping, and adds a contrastive user personalized adaptation loss aligning a user's representations across domains.
Key results
- On ToysG-VideoG, D2C2R reaches MRR 4.68 on ToysG versus 4.52 for the strongest baseline CDRIB, and 6.35 on VideoG versus 6.27 for UniCDR.
- Ablation on Music/Movie/Phones/Ele shows removing CDKF or UPA lowers NDCG and HR relative to the full D2C2R model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CDKF (information enhancement and heuristic cross-domain fusion) (ablation) | D2C2R · ToysG | ndcg: 4.03 -> 4.68 | Helped |
| add UPA (user personalized adaptation) (ablation) | D2C2R · ToysG | ndcg: 3.90 -> 4.68 | Helped |
| add both CDKF and UPA (ablation) | D2C2R · ToysG | ndcg: 2.98 -> 4.68 | Helped |
| add CDKF (information enhancement and heuristic cross-domain fusion) (ablation) | D2C2R · VideoG | ndcg: 6.42 -> 6.93 | Helped |
| add UPA (user personalized adaptation) (ablation) | D2C2R · Music | hr: 9.29 -> 10.13 | Helped |
| use D2C2R instead of CDRIB | ToysG · ToysG | MRR: 4.52±0.22 -> 4.68±0.31 | Helped |
| use D2C2R instead of UniCDR | VideoG · VideoG | MRR: 6.27±0.19 -> 6.35±0.36 | Helped |
Limitations
- Gains over the strongest baselines are small on some metrics (e.g., VideoG MRR 6.35 vs 6.27 for UniCDR).
- Performance declines for some source-interaction ranges in Phones-Ele, which the authors address only partially with UPA.
Experimental setup
Datasets: Amazon ToysG-VideoG Amazon Music-Movie Amazon Phones-Ele
Benchmarks: Amazon ToysG-VideoG Amazon Music-Movie Amazon Phones-Ele
Metrics: MRR NDCG@5 NDCG@10 HR@1 HR@5 HR@10
Baselines: BPRMF NGCF EMCDR DCDCSR SSCDR TMCDR SA-VAE CDRIB UniCDR
Builds on: VGAE EMCDR CDRIB UniCDR
Closest papers by meaning
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.91)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.89)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.88)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.86)
- One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation (AAAI 2025 · similarity 0.85)
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.85)
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.