UGR generates Taobao's organic, ad and cross-business candidates in one pass, lifting online transaction volume 3.29%
Synopsis
Ranking in Taobao's Guess You Like service no longer generates a list over a truncated single-business candidate set and then hands it to cross-business blending. UGR generates one list over roughly a thousand organic, ad and other-business candidates per request, models historical ad exposure separately during training, and adds current ad value at serving. In an online A/B test on Guess You Like, UGR raised Core IPV by 1.76%, Transaction Volume by 3.29% and GMV by 2.10% over the production baseline.
Interpretation
A ranking list is now generated in one pass over the full request-level candidate set, so organic items, ads and items from other businesses compete for display positions along a single decoding path instead of being generated per business and blended afterward. Earlier generative ranking methods typically worked on truncated, single-business candidate sets, and the generated list was then blended with items from other businesses; blending changes item positions and list context, creating a discrepancy between the generated list and the final exposure list. On Taobao's Guess You Like, UGR-Single+Mix, which restricts UGR to organic-only lists and leaves blending to the existing mixrank, improved Core IPV, Transaction Volume and GMV over the same production baseline by 1.13%, 2.84% and 1.38%, while UGR improved them by 1.76%, 3.29% and 2.10%, and UGR significantly outperformed the staged variant in the direct comparison.
Beyond imitating logged exposure sequences, generation is also constrained by post-exposure click and transaction feedback that shapes list order. Learning only from exposure sequences reproduces the selection left by the existing ranking policy and cannot fully use how users responded after seeing items. In the Scenario 1 ablation, removing PermPO lowered all six ranking metrics, with IPV NDCG falling from 0.3666 to 0.3453; removing WSFT gave IPV NDCG of 0.3539, against 0.3382 for NTP-only training.
Historical ad exposure effects are modeled in a tower that exists only during training; the tower is removed at serving and current ad value is added back to the recommendation score. Fitting exposure lists directly embeds the selection effects of historical ad-serving policies into recommendation scores that are later combined with current ad value for allocation and pricing. Without the tower, generator logits correlate positively with ad eCPM overall; with the tower, the correlation between generator logits and eCPM becomes negative while the tower adjustment correlates positively with eCPM, indicating that eCPM-related variation sits mainly in the tower.
Perspective
The demonstrated scope is Taobao's homepage Guess You Like recommendation service: offline experiments use two industrial datasets built from 30 consecutive days of that service's recommendation logs, and the online A/B test ran with the UGR group covering at least 5% of traffic for more than 14 days while the downstream evaluator and the rest of the serving pipeline stayed unchanged.
Online gains are reported as percentages relative to the production baseline and come from the specific Guess You Like service; offline, UGR does not lead on every metric, since UniRank achieves the highest Hitrate in both scenarios while UGR leads on feedback-based ranking quality.