Research map / Label-efficient and robust learning / Domain adaptation and generalization
Towards Unbiased Information Extraction and Adaptation in 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 UIEA, a multi-target cross-domain recommendation method that extracts domain-invariant global user embeddings by optimizing worst-case (group DRO) domain loss and adapts them to target domains via a user-item attention module.
Problem
Multi-target cross-domain recommendation under data sparsity: average-loss knowledge extraction absorbs biased information from data-rich domains and hurts data-scarce domains, and existing adaptation ignores item-side information.
Approach
Replace average-case loss in the extraction stage with a group-DRO min-max objective (joint model and domain-weight updates) to learn global user embeddings, and adapt them to each target domain with a user-item attention module that uses averaged embeddings of historically interacted items as the query.
Key results
- UIEA improves MRR@10 by 15.23% on Amazon Movies and by 22.32% on Douban Music over the best cross-domain baseline.
- In the worst-case domain on Amazon, UIEA improves MRR@10, HR@10 and NDCG@10 by 14.37%, 5.47% and 12.29%.
- Ablation on Amazon shows UIE+MLP and AVG+UIA each improve over AVG+MLP, with the full UIEA best in all domains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add worst-case UIE and user-item attention UIA | Cross-domain recommendation · Amazon Movies | HR@10: UIEA: 62.56 ± 0.17; BPRMF: 56.31 ± 0.04 | Helped |
| Add worst-case UIE and user-item attention UIA | Cross-domain recommendation · Douban Music | HR@10: UIEA: 40.60 ± 0.57; CMF: 36.07 ± 1.74 | Helped |
| Add worst-case UIE and user-item attention UIA | Cross-domain recommendation · IE KR | HR@10: UIEA: 53.86 ± 0.18; HeroGraph: 52.37 ± 0.94 | Helped |
| Add fine-tuning loss to reconstruction-only extraction (ablation) | UIEA · Amazon Books | HR@10: 47.15 ± 0.15 -> 48.48 ± 0.20 | Helped |
| Replace average-case with worst-case extraction (ablation) | UIEA with MLP adaptation · Amazon Movies | HR@10: 55.35 ± 0.28 -> 59.98 ± 0.05 | Helped |
| Replace average-case with worst-case extraction (ablation) | UIEA with user-item attention · Amazon Elec | HR@10: 35.40 ± 0.19 -> 41.83 ± 0.31 | Helped |
| Add user-item attention instead of MLP adaptation (ablation) | UIEA with UIE extraction · Amazon Movies | HR@10: 59.98 ± 0.05 -> 62.56 ± 0.17 | Helped |
| Add user-item attention instead of MLP adaptation (ablation) | UIEA with UIE extraction · Amazon Elec | HR@10: 37.52 ± 0.18 -> 41.83 ± 0.31 | Helped |
| Use reconstruction-only extraction and MLP adaptation rather than full UIEA (ablation) | UIEA · Amazon Movies | HR@10: 62.56 ± 0.17 -> 54.02 ± 0.17 | Hurt |
Surprising findings
- Rec+MLP improves over single-domain BPRMF only in the Books domain and degrades in Movies and Elec, indicating overfitting to data-rich domains.
Limitations
- Evaluation limited to three datasets with three domains each, all using a sampled leave-one-out protocol with 99 negatives.
- Method is evaluated only against recommendation baselines and only on ranking metrics; no analysis of computational cost is given.
Experimental setup
Datasets: Amazon (Books, Movies, Elec) Douban (Books, Movies, Music) IE (BR, KR, US)
Benchmarks: Amazon Douban IE
Metrics: MRR@10 HR@10 NDCG@10
Baselines: BPRMF NeuMF CMF HeroGraph GA-MTCDR EDDA CAT-ART
Builds on: Bayesian personalized ranking (BPR-MF) group distributionally robust optimization scaled dot-product attention
Closest papers by meaning
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.88)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.87)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.86)
- Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation (ICML 2024 · similarity 0.86)
- One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation (AAAI 2025 · similarity 0.86)
- 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.