OneTrans-GSR unifies reranking and mix-ranking in recommender systems
Original title: OneTrans-GSR: A Generative Slate Ranking Paradigm Unifying Reranking and Mix-Ranking with One Autoregressive Transformer
OneTrans-GSR generates slates end-to-end with reinforcement learning and improves GMV/u by 2.63% while cutting latency by 13%.
Industrial recommender systems use separate stages for candidate scoring, reranking, and mix-ranking with hand-crafted rules. OneTrans-GSR reformulates these as a unified problem solved by a single autoregressive transformer policy. A reward model guides training toward slate-level objectives with business constraints enforced at decoding. Offline and online results show consistent gains as compute increases.
- OneTrans-GSR mixes heterogeneous content types in one slate
- Improves GMV/u by 2.63% in A/B tests
- Reduces end-to-end latency by 13%
- Reward model evaluations guide reinforcement learning
- Policy extends deployed OneTrans backbone
From the original article
1 min read at huggingface.co
Industrial recommender systems allocate substantial computation to pointwise candidate scoring, yet separate reranking and mix-ranking stages still assemble the final exposure slate using hand-crafted rules and specialized low-capacity models. Mix-ranking combines product cards, short videos, livestreams, and advertising cards supplied by distinct upstream business channels.
By Feng Cheng, Yu-Hsuan Huang, Fengyi Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.