Research map / Probabilistic and scientific modeling / Molecules and proteins
Parallel Ranking of Ads and Creatives in Real-Time Advertising Systems
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 Peri-CR, a parallel architecture that ranks ads and creatives simultaneously in online advertising, plus JAC, an offline joint CTR model, and NSCTR, a normalized offline creative-ranking metric.
Problem
Creative ranking in real-time ad systems is usually done serially before or after ad ranking, which limits creative modeling sophistication and adds latency; offline metrics for creative ranking (AUC, GAUC, sCTR) are also unreliable.
Approach
Decouple creative ranking from ad ranking so both run in parallel online, and train an offline joint cascaded model (JAC) where the creative ranker consumes the ad model's quantized CTR, with gradients copied back to the ad model; a new metric NSCTR normalizes simulated CTR by ad distribution.
Key results
- Peri-CR+ achieved +10.12% CTR and +7.67% RPM over no-CR in 5-day online A/B test with response time of 90ms, matching no-CR, while Pre-CR (CACS) took 107ms.
- JAC reached the highest NSCTR of 0.2427 (+7.0% over VAM-HBM), and NSCTR correlated with online CTR lift at Pearson 0.988 across 6 upgrades.
- Offline JAC raised ad-ranking AUC from 74.58% to 74.87% and GAUC from 71.83% to 72.14% over the AR baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use parallel ad and creative ranking instead of Pre-CR | Peri-CR+ · Online A/B test | CTR: +1.58% | Helped |
| Use parallel ad and creative ranking instead of Pre-CR | Peri-CR+ · Online A/B test | RPM: +0.86% | Helped |
| Use parallel ranking rather than serial Post-CR or Pre-CR | Peri-CR+ · Online A/B test | RT: 90ms; Post-CR 94ms; Pre-CR 107ms | Helped |
| Add JAC guidance to the creative ranker (ablation) | CR+ · Creative ranking test set | NSCTR: 0.2392 -> 0.2423 | Helped |
| Add JAC guidance to the creative ranker (ablation) | CR+ · Creative ranking test set | sCTR: 0.2559 -> 0.2492 | Hurt |
| Add JAC guidance to the ad ranker (ablation) | AR+ · Ad ranking test set | AUC: 74.58% -> 74.70%(+0.12%) | Helped |
| Add JAC guidance to the ad ranker (ablation) | AR+ · Ad ranking test set | GAUC: 71.83% -> 71.98%(+0.15%) | Helped |
| Jointly model ad and creative ranking | JAC · Creative ranking test set | NSCTR: 0.2427(+7.0%) | Helped |
| Jointly model ad and creative ranking | JAC · Creative ranking test set | sCTR: 0.2478(-2.0%) | Mixed |
| Jointly model ads and creatives | ACR · Ad ranking test set | AUC: 74.87%(+0.29%) | Helped |
| Jointly model ads and creatives | ACR · Ad ranking test set | GAUC: 72.14%(+0.31%) | Helped |
| Add JAC guidance to Peri-CR (ablation) | Peri-CR+ · Online A/B test | CTR: +0.92% | Helped |
Surprising findings
- Pearson correlation of GAUC with online CTR lift was -0.152, indicating GAUC is unreliable for creative ranking evaluation.
- The sCTR metric favored the offline JAC less (JAC had -2.0% sCTR vs VAM-HBM) even though it had the best NSCTR.
Limitations
- A large joint ranker for ads and creatives could not be deployed online due to latency constraints.
- NSCTR only approximates the true sample distribution and does not fully remove selection bias.
- Creatives are produced offline and not generated in real time for each user.
Experimental setup
Datasets: proprietary JD.com ad click logs
Benchmarks: proprietary JD.com ad log test set
Metrics: AUC GAUC sCTR NSCTR CTR RPM response time
Baselines: no-CR Post-CR Pre-CR (CACS) VAM-HBM DSSM two-tower
Builds on: CACS DSSM Deep & Cross Network bandit algorithms
Closest papers by meaning
- GenAuction: A Generative Auction for Online Advertising (AAAI 2025 · similarity 0.84)
- Learning Personalized Ad Impact via Contextual Reinforcement Learning under Delayed Rewards (NeurIPS 2025 · similarity 0.83)
- Beyond Last-Click: An Optimal Mechanism for Ad Attribution (NeurIPS 2025 · similarity 0.82)
- LLM-Auction: Generative Auction towards LLM-Native Advertising (NeurIPS 2026 · similarity 0.82)
- AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.82)
- Optimal Auction Design in the Joint Advertising (ICML 2025 · 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.