Generating mixed recommendation lists improved tested outcomes
UGR: Unified Generative Ranking for Heterogeneous E-commerce Recommendation · arXiv:2610.39426
In plain words
This paper introduces Unified Generative Ranking, a method for e-commerce recommendations that makes one ordered list from products, ads and candidates from other businesses.
It uses the request and the items already placed to choose what comes next, instead of ranking business groups separately and blending them later.
Across two offline scenarios, it ranked items better by observed engagement and transactions than the tested alternatives, though another method matched logged lists more often. A live test on Taobao’s Guess You Like service also improved visits, orders and paid-order value over production.
The research
When a recommendation request contains products, ads and items from other businesses, separate ranking followed by blending can change item positions and the surrounding list.
It remains unclear whether the full mixed set can be ranked in one generation process while still using observed user feedback and current ad value under existing serving rules.
UGR generates one ranked list from the full mixed candidate set, trains on exposure and feedback, and leaves final exposure selection to the existing evaluator.
In two offline scenarios, UGR ranked highest on six measures comparing lists with observed engagement and transaction feedback. UniRank had the most overlap with logged exposures, showing that agreement with past displays differed from ranking by feedback.
On Taobao’s Guess You Like service, an online test against production found higher core item-page visits (+1.76%), completed orders (+3.29%) and paid-order value (+2.10%). The existing evaluator and remaining serving pipeline were unchanged.
The evidence covers two offline scenarios and an online test on the same Taobao service. Offline rankings against logged feedback are not live outcomes, and the paper reports no results from other services or platforms.
What the evidence shows
- Rank the mixed list
Unified Generative Ranking (UGR) is a framework for producing recommendation lists from the full set of candidates in a request, including regular items, ads and candidates from other businesses. It addresses a specific mismatch: separately generated lists can change when they are blended afterward.
UGR generates the mixed list directly, while the existing downstream evaluator still selects the final list for exposure.
- Use feedback and ad value
UGR learns from logged exposure lists and from observed engagement and transaction feedback. It weights exposed items by feedback and compares alternative orders of the same items, encouraging higher-feedback items to appear earlier.
It also models historical ad exposure effects with a training-only tower, then incorporates current ad value during serving under the existing allocation and pricing rules.
- Separate exposure from feedback
In two offline scenarios from Taobao’s Guess You Like service, UGR ranked highest on six measures of ranking against observed engagement and transaction feedback. UniRank had the highest Hitrate, which measures overlap with logged exposure lists.
The contrast shows that matching past displays and ranking by observed feedback can favor different methods.
- Check the live service
An online test on the same service found higher core item-page visits, completed orders and paid-order value for UGR than for production. The downstream evaluator and remaining serving pipeline stayed unchanged between groups.
The reported gains therefore support the ranking system on this service under the tested conditions, rather than establishing results for other services.
- Test generation and training
A second online comparison found that UGR outperformed a staged version that generated organic-only lists before cross-business blending. In a Scenario 1 training ablation, full UGR also exceeded next-token-only training on feedback-ranking measures, with comparable exposure overlap.
That ablation changed several training components together, so it does not identify each component’s separate effect.
Deep
- Original paper authors
- Zonghe Shao, Huabin Chen, Sitong Lu, Fangmei Zhu, Yingcai Ma
- Submitted to arXiv
- License
- CC BY 4.0