Research map / Label-efficient and robust learning / Recognition and long-tail
AT4CTR: Auxiliary Match Tasks for Enhancing Click-Through Rate Prediction
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 AT4CTR, adding two InfoNCE-based auxiliary match tasks (user-item match and next item prediction) to CTR models to alleviate data sparsity caused by negative down-sampling.
Problem
CTR prediction suffers from data sparsity, which negative sample down-sampling in industrial systems worsens, degrading performance; prior auxiliary methods either use weak negative-sampling losses or are costly.
Approach
Two auxiliary match tasks with InfoNCE loss: User-Item Match pulls user and item representations of clicked positives together using in-batch negatives, and Next Item Prediction uses causal self-attention over click history to predict the next item; both add training signal without extra inference cost or parameters.
Key results
- On the Taobao dataset AT4CTR reaches AUC 0.9535 and Logloss 0.1345, beating the best baseline DeepMCP (0.9514 / 0.1376).
- On the Industry dataset AT4CTR reaches AUC 0.7441 vs 0.7428 for the best baseline CL4CTR.
- Online A/B test on 10% traffic for one month gave 1.27% relative lift in Revenue Per Search and 7.21% relative ROI increase over DBPMaN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UIM and NIP auxiliary match tasks | DBPMaN · Industry | AUC: 0.7426 -> 0.7441 | Helped |
| add UIM and NIP auxiliary match tasks | DBPMaN · Industry | Logloss: 0.1905 -> 0.1902 | Helped |
| add UIM and NIP auxiliary match tasks | DBPMaN · Taobao | AUC: 0.9509 -> 0.9535 | Helped |
| add UIM and NIP auxiliary match tasks | DBPMaN · Taobao | Logloss: 0.1382 -> 0.1345 | Helped |
| add UIM (ablation) | DBPMaN · Industry | AUC: 0.7426 -> 0.7438 | Helped |
| add UIM (ablation) | DBPMaN · Industry | Logloss: 0.1905 -> 0.1903 | Helped |
| add UIM (ablation) | DBPMaN · Taobao | AUC: 0.9509 -> 0.9529 | Helped |
| add UIM (ablation) | DBPMaN · Taobao | Logloss: 0.1382 -> 0.1360 | Helped |
| add NIP to DBPMaN + UIM (ablation) | DBPMaN + UIM · Industry | AUC: 0.7438 -> 0.7441 | Helped |
| add NIP to DBPMaN + UIM (ablation) | DBPMaN + UIM · Industry | Logloss: 0.1903 -> 0.1902 | Helped |
| add NIP to DBPMaN + UIM (ablation) | DBPMaN + UIM · Taobao | AUC: 0.9529 -> 0.9535 | Helped |
| add NIP to DBPMaN + UIM (ablation) | DBPMaN + UIM · Taobao | Logloss: 0.1360 -> 0.1345 | Helped |
Surprising findings
- Performance peaks at negative down-sampling ratio 0.3 rather than 1.0 on the industry dataset, which the authors leave unexplained.
- Existing auxiliary methods such as CL4CTR and DMR show almost no improvement over baselines.
Limitations
- NIP gains are limited because behavior sequence items carry few features (only item_id, category_id) due to storage cost.
- NIP is sensitive to loss weight and must be kept at a small magnitude.
- Industry dataset is proprietary, limiting reproducibility.
Experimental setup
Datasets: Taobao dataset Meituan industry search advertising dataset
Benchmarks: Taobao Meituan industry dataset
Metrics: AUC Logloss revenue per search return on investment
Baselines: FM WideDeep DeepFM DIN DIEN DBPMaN DeepMCP DMR CL4CTR
Builds on: DBPMaN DIN DIEN InfoNCE CL4CTR
Closest papers by meaning
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.87)
- RecFlow: An Industrial Full Flow Recommendation Dataset (ICLR 2025 · similarity 0.86)
- From Feature Interaction to Feature Generation: A Generative Paradigm of CTR Prediction Models (ICML 2025 · similarity 0.86)
- LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.84)
- Entire-Space Variational Information Exploitation for Post-Click Conversion Rate Prediction (AAAI 2025 · similarity 0.84)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · 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.