Research map / Label-efficient and robust learning / Domain adaptation and generalization
D3: A Methodological Exploration of Domain Division, Modeling, and Balance in Multi-Domain Recommendations
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 D3, a plug-in framework for multi-domain CTR recommendation that learns domain division via attention-based feature selection, fuses shared and specific weights with a gate, and reweights losses by domain complexity and specificity.
Problem
Multi-domain recommendation relies on manual domain partitioning, overlooks inter-domain relationships, and suffers from training inconsistency across domains during joint optimization.
Approach
Instance-level attention selects domain-sensitive features for dynamic domain division, a gate fuses shared and domain-specific weights, and loss weights are derived from attention entropy (complexity) and gate value (specificity).
Key results
- Plugging D3 into Shared-Bottom, MMOE, M2M, ADI, STAR and SAR-Net improves AUC on all three datasets; on Aliccp, ADI+D3 raises S#2 AUC from 0.5995 to 0.6029.
- On Aliccp with ADI backbone, D3 (0.6261/0.6029/0.6230 AUC) outperforms M2M-WG, PEPNet-WG, DWA and DT weighting alternatives.
- Ablation on Aliccp shows each component (DSFS, DF, DBO) adds gains that accumulate; loss adaptation is most effective when applied between 50% and 75% of training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add D3 to ADI | ADI · Aliccp S#1 | AUC: 0.6214 -> 0.6261 | Helped |
| Add D3 to ADI | ADI · Movielens-1M S#1 | AUC: 0.7741 -> 0.7830 | Helped |
| Add D3 to ADI | ADI · ADC S#1 | AUC: 0.5775 -> 0.5840 | Helped |
| Add DSFS (domain division) (ablation) | BM · Aliccp S#1, S#2, and S#3 | AUC: 0.6214 -> 0.6240; 0.5995 -> 0.6000; 0.6175 -> 0.6209 | Helped |
| Add DF (domain modeling) after DSFS (ablation) | BM+D1 · Aliccp S#1, S#2, and S#3 | AUC: 0.6240 -> 0.6248; 0.6000 -> 0.6018; 0.6209 -> 0.6221 | Helped |
| Add DBO (domain balance optimization) after DSFS and DF (ablation) | BM+D2 · Aliccp S#1, S#2, and S#3 | AUC: 0.6248 -> 0.6261; 0.6018 -> 0.6029; 0.6221 -> 0.6230 | Helped |
| Use D3 domain-sensitive feature selection and fusion instead of M2M-WG or PEPNet-WG | ADI backbone on Aliccp · Aliccp S#1, S#2, and S#3 | AUC: 0.6230, 0.6233 -> 0.6248; 0.5978, 0.5996 -> 0.6018; 0.6204, 0.6203 -> 0.6221 | Mixed |
| Use D3 loss adaptation instead of DWA or DT | ADI backbone with DSFS and DF on Aliccp · Aliccp S#1, S#2, and S#3 | AUC: 0.6156, 0.6229 -> 0.6261; 0.5942, 0.5952 -> 0.6029; 0.6124, 0.6122 -> 0.6230 | Helped |
Surprising findings
- Applying loss adaptation (DWA, DT) on top of D2 reduced AUC relative to D2 alone on Aliccp, suggesting generic multi-task loss weighting can hurt multi-domain training.
Limitations
- Loss adaptation timing is sensitive; it helps only when introduced mid-training.
- Evaluation limited to CTR AUC/Logloss on three public datasets; online gains described only qualitatively.
Experimental setup
Datasets: Aliccp MovieLens-1M ADC
Benchmarks: Aliccp MovieLens-1M ADC
Metrics: AUC Logloss
Baselines: Shared-Bottom MMOE M2M ADI STAR SAR-Net M2M-WG PEPNet-WG DWA DT
Builds on: STAR M2M PEPNet DWA Dynamic Weight Adaptation
Closest papers by meaning
- Multi-Domain Recommendation to Attract Users via Domain Preference Modeling (AAAI 2024 · similarity 0.87)
- One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation (AAAI 2025 · similarity 0.86)
- RecFlow: An Industrial Full Flow Recommendation Dataset (ICLR 2025 · similarity 0.85)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.85)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.84)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.84)
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.