Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Domain Deep Learning from a Multi-View Perspective for Cross-Border E-commerce Search
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 DAMO, a domain-number-invariant multi-view CTR/CVR model with a lightweight view-attention estimator, evaluated on an industrial cross-border e-commerce dataset, Ali-CCP, and online A/B testing.
Problem
Multi-domain CTR/CVR prediction for cross-border e-commerce search suffers from poor scalability and efficiency as the number of domains (countries) grows, since domain-specific parameters scale linearly with domain count.
Approach
Maps multi-domain features into a low-rank multi-view space (view-invariant projection features plus view-specific features) with domain-aware gating and a dictionary-learning-based view-attention estimator that weights view-guided losses, so model size does not grow with the number of domains.
Key results
- On the industrial dataset (top-5 to top-50 domains), DAMO averages CTR AUC 0.7350, CVR AUC 0.8233, CTCVR AUC 0.8794, outperforming MMoE, CGC, STAR, M2M and MSSM.
- DAMO's computational cost stays constant as domains increase, e.g., 6.28B FLOPs and 3.19M parameters, with the lowest cost among compared methods.
- Online A/B test over two weeks with 4% traffic per group improved CTR by 1.21%, CVR by 1.76%, and GMV by 1.66% over the existing CGC-based model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Correlated Multi-view Extractor instead of CGC shared and specific experts (ablation) | DAMO · Industrial dataset, 5 clusters | CTR AUC: 0.7692 ± 4.3e-5 -> 0.7698 ± 1.1e-4 | Helped |
| Add Domain-aware Multi-view Predictor instead of one-view gating network followed by MLP (ablation) | DAMO · Industrial dataset, 5 clusters | CTCVR GAUC: 0.6994 ± 9.3e-4 -> 0.7023 ± 3.9e-4 | Helped |
| Add Lightweight View-attention Estimator instead of fixed βv=0.2 (ablation) | DAMO · Industrial dataset, 5 clusters | CTCVR AUC: 0.9155 ± 7.3e-4 -> 0.9174 ± 2.6e-4 | Helped |
| Use DAMO for ranking-model CTR/CVR predictions rather than the previous model | Online e-commerce ranking model · Online A/B test | CTR: +1.21% | Helped |
Surprising findings
- M2M is second-best in accuracy on the industrial dataset but performs poorly on Ali-CCP, and MSSM underperforms on several metrics.
Limitations
- Efficiency claims rely on FLOPs/parameter counts; the industrial dataset is proprietary and results are not independently reproducible.
- Ali-CCP has only three domains, so domain scalability is only tested on the industrial dataset.
- The reported accuracy gains on offline metrics are small (around 0.05% AUC).
Experimental setup
Datasets: Ali-CCP industrial cross-border e-commerce click/purchase logs
Benchmarks: Ali-CCP industrial cross-border e-commerce dataset
Metrics: AUC GAUC FLOPs parameters GMV CTR CVR
Baselines: MMoE CGC STAR M2M MSSM
Builds on: MMoE CGC STAR deep dictionary learning GAUC
Closest papers by meaning
- GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction (ICML 2026 · similarity 0.82)
- TIGER-FG: Text-Guided Implicit Fine-Grained Grounding for E-commerce Retrieval (NeurIPS 2026 · similarity 0.82)
- D3: A Methodological Exploration of Domain Division, Modeling, and Balance in Multi-Domain Recommendations (AAAI 2024 · similarity 0.82)
- Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.82)
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.81)
- eCeLLM: Generalizing Large Language Models for E-commerce from Large-scale, High-quality Instruction Data (ICML 2024 · similarity 0.81)
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.