Research map / Label-efficient and robust learning / Domain adaptation and generalization
Multi-Domain Recommendation to Attract Users via Domain Preference Modeling
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.
Introduces the MDRAU task (recommending from multiple unseen domains) and DRIP, a masked-domain-modeling framework with a self-attention multi-domain encoder that jointly models domain-level and item-level preferences.
Problem
Recommending items from multiple unseen service domains that users have not interacted with yet, using knowledge from their seen domains, when the number of seen-unseen domain combinations is large and users have varying domain-level preferences.
Approach
Decompose user preference into domain-level (softmax over domains via a [S] token) and item-level (in-domain similarity) components, and train a multi-domain transformer encoder by randomly masking seen-domain user embeddings and predicting their interactions, with adaptive masking weighted by predicted domain preference.
Key results
- On MDRAU-MT P1, DRIP reaches R@20 0.0545 and N@20 0.0556, improving over the best baseline UniCDR (0.0434 / 0.0446) by about 25.6% and 24.6%.
- On MDRAU-ST, DRIP achieves the best R@20 in every target domain across P1 and P2, e.g., Book R@20 0.0403 vs 0.0359 for UniCDR.
- Single-domain learning drops MDRAU-MT P1 R@20 to 0.0281, and removing adaptive masking lowers R@20 to 0.0513 versus 0.0545 for full DRIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use unified DRIP rather than UniCDR | MDRAU recommendation · MDRAU-MT P1 | R@20: 0.0434 -> 0.0545 | Helped |
| use unified DRIP rather than UniCDR | MDRAU recommendation · MDRAU-MT P2 | N@20: 0.0574 -> 0.0773 | Helped |
| replace unified joint training with single-domain learning (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0281 | Hurt |
| replace unified joint training with many-to-one learning and post-processing A (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0432 | Hurt |
| replace unified joint training with many-to-one learning and post-processing B (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0409 | Hurt |
| replace personalized domain preference with a uniform distribution (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0122 | Hurt |
| replace personalized domain preference with domain activeness distribution (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0409 | Hurt |
| remove adaptive masking and use random masking at the same ratio (ablation) | DRIP · MDRAU-MT P1 | R@20: 0.0545 -> 0.0513 | Hurt |
Surprising findings
- BPRMF, which treats all domains as one, performs poorly on MDRAU-ST but is comparable to state-of-the-art methods on MDRAU-MT, attributed to implicit domain-level preference capture.
- Fixed domain-preference distributions (uniform or domain activeness) yield substantially worse performance than personalized domain-level preference.
Limitations
- Evaluation uses only Amazon datasets with simulated platform scenarios by hiding domains of overlapping users.
- Post-processing is needed for baselines to produce unified multi-domain lists, which may not be optimal.
- Performance drops sharply when the masking probability exceeds 0.7, indicating sensitivity to masking hyperparameters.
Experimental setup
Datasets: Amazon
Benchmarks: Amazon P1 (Book, Movie, CD, Game) Amazon P2 (Home, Health, Grocery, Tools)
Metrics: Recall@K NDCG@K KL divergence (KLD@K)
Baselines: BPRMF MMOE PLE EMCDR+ PTUPCDR+ CAT-ART+ UniCDR
Builds on: BPRMF EMCDR PTUPCDR UniCDR CAT-ART
Closest papers by meaning
- D3: A Methodological Exploration of Domain Division, Modeling, and Balance in Multi-Domain Recommendations (AAAI 2024 · similarity 0.87)
- One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation (AAAI 2025 · similarity 0.85)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.84)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.84)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.83)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.83)
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.