Research map / Label-efficient and robust learning / Recognition and long-tail
One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation
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 AREAD, a hierarchical expert network with iterative lottery-ticket-style mask pruning per domain and popularity-based counterfactual augmentation for minor domains, for multi-domain recommendation across 20+ domains.
Problem
Multi-domain recommendation becomes hard when there are dozens of domains: parameter count grows with per-domain towers, cross-domain transfer patterns are complex, and minor domains are data-sparse and dominated by major ones.
Approach
A hierarchical expert integration module (few experts at lower layers, more at higher layers) with per-domain hierarchical expert masks found by iterative pruning (lottery-ticket style) and counterfactual augmentation that transfers unpopular-item interactions from major to minor domains under explain-away and cross-domain interest-consistency assumptions.
Key results
- On Amazon, AREAD improves overall AUC by 6.7 per-thousand over baselines (0.7120 vs. 0.7053 best baseline) and on AliCCP improves AUC to 0.6122 vs. 0.6091 (MMoE).
- On AliCCP, AREAD raises Minor10AUC to 0.6200 and Minor5AUC to 0.6264, versus 0.6151 and 0.6131 for the best baseline ADL.
- Mask overlap analysis shows Home & Kitchen and Appliances share experts (OR 0.405 vs. random 0.25) while Home & Kitchen and Movies & TV overlap little (OR 0.065).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AREAD instead of the strongest baseline | Multi-domain recommendation · Amazon | AUC: +6.7‰ | Helped |
| Use AREAD instead of the strongest mixed-domain baseline | Multi-domain recommendation · Amazon | DomainAUC: +2‰ | Helped |
| Use AREAD instead of the strongest baseline | Multi-domain recommendation · AliCCP | AUC: +3‰ | Helped |
| Use AREAD instead of the strongest mixed-domain baseline | Multi-domain recommendation · AliCCP | DomainAUC: +8.4‰ | Helped |
| Add the HEI module without expert selection masks (ablation) | MMoE base recommender · Amazon and AliCCP | AUC | Hurt |
| Add HEMP domain-specific expert masks (ablation) | AREAD with HEI · Amazon and AliCCP | AUC | Helped |
| Add HEMP domain-specific expert masks (ablation) | AREAD with HEI · Amazon and AliCCP | DomainAUC | Helped |
| Add popularity-based counterfactual augmentation (ablation) | AREAD with HEI and HEMP · Amazon and AliCCP | Minor10AUC | Helped |
| Increase HEI depth from 2 to 3 layers (ablation) | AREAD · Amazon | performance | Helped |
| Increase HEI depth from 3 to 4 layers (ablation) | AREAD · Amazon | performance | Helped |
| Increase first-layer experts from fewer experts to a larger number (ablation) | AREAD · Amazon | Minor-domain performance | Helped |
| Increase augmentation ratio from 0.05 to 0.1 (ablation) | AREAD · Amazon | Minor10AUC | Helped |
Surprising findings
- Adding the hierarchical expert integration module alone slightly degrades performance relative to the MMoE base recommender, per the ablation.
- On Amazon, the Isolated per-domain model achieves higher DomainAUC than AREAD, though AREAD is far cheaper to maintain.
Limitations
- Counterfactual augmentation transfers items across domains and can introduce noise; large augmentation ratios significantly hurt performance.
- Does not surpass the Isolated per-domain method on Amazon DomainAUC.
- Relies on clustering-free but compute-heavy iterative mask search (Z candidate masks per domain, periodic pruning).
Experimental setup
Datasets: Amazon AliCCP
Benchmarks: Amazon AliCCP
Metrics: AUC DomainAUC Major5AUC Minor10AUC Minor5AUC
Baselines: DeepFM DCN AutoInt DCNv2 EPNet-S Isolated MMoE PLE STAR AdaSparse HiNet EPNet
Builds on: MMoE Lottery Ticket Hypothesis AESM^2 DCN
Closest papers by meaning
- D3: A Methodological Exploration of Domain Division, Modeling, and Balance in Multi-Domain Recommendations (AAAI 2024 · similarity 0.86)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.86)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.85)
- Multi-Domain Recommendation to Attract Users via Domain Preference Modeling (AAAI 2024 · similarity 0.85)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.85)
- LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (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.