OneTrans-V2: A Unified Transformer for Industrial Recommendations

Paper: OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender

Authors: Hannan Cao, Jun Guo, Haolei Pei, Zhaoqi Zhang, Tianyu Wang, Ziyang Wang, Youchen Sun, Yue Xue, Yucheng Mao, Lintao Yan, Yufei Feng, Shaowei Liu, Rongkun Xing, Feiling Gong, Xinyu Chenli, Cong Xu, Mingge Zhang, Yunjia Zhu, Yajing Zhang, Pengfei Ren, Yue Lin

Executive Summary

The research introduces OneTrans-V2, an advanced recommendation system designed to improve the efficiency and effectiveness of industrial recommendation engines. Traditional systems operate as a series of separate stages—retrieval, pre-ranking, and fine-ranking—each requiring independent models. This separation leads to inefficiencies, such as repeated processing of user data and increased engineering efforts.

OneTrans-V2 addresses these challenges by unifying all three stages into a single model. This integration allows the system to encode user behavior just once, creating a shared context that enhances the overall performance. By incorporating features unique to each stage while still maintaining a cohesive architecture, OneTrans-V2 enables the stages to work together more effectively. This joint training approach fosters knowledge sharing between the fine-ranking and pre-ranking processes, potentially improving the quality of recommendations.

A key innovation in OneTrans-V2 is the use of sparse mixture-of-experts (MoE) to scale the model's capacity. MoE allows for greater computational efficiency by activating only a subset of the model’s parameters during processing, thus optimizing performance without incurring additional costs. The research also introduces a novel training method called Decision-Conditioned Generative Retrieval (DCGR). This method predicts user interactions based on prior behavior, generating tailored item recommendations that align more closely with business goals.

OneTrans-V2 also implements Sequence-Native Training (SNT), which organizes the training process around each user's complete behavioral history. This approach allows the model to better understand user preferences over time and amortizes the computational costs of encoding user behavior across multiple interactions.

In practical applications, OneTrans-V2 has been deployed in a large-scale industrial setting. The results show a significant improvement in gross merchandise value (GMV) by 9.74%, indicating that the enhanced recommendation capabilities can lead to increased sales and revenue. Additionally, the system achieves a throughput that is 3.2 times greater than the traditional cascade model, all while utilizing the same hardware resources. This improvement suggests that companies could achieve higher efficiency and better performance without needing to invest in additional infrastructure.

Overall, OneTrans-V2 represents a step forward in the development of recommendation systems, with potential implications for businesses seeking to enhance user engagement and drive sales. The research highlights the benefits of unified modeling and innovative training techniques, paving the way for more effective and efficient recommendation solutions in various industrial contexts.

Academic Abstract

Industrial recommendation systems typically operate as a \emph{cascade} of retrieval, pre-rank, and fine-rank, but these stages are usually trained and served as separate models, causing repeated user-sequence encoding, isolated optimization, and duplicated engineering effort. Building on OneTrans' model-level unification, we present OneTrans-V2, one Transformer that unifies the entire cascade. It encodes the user behavior sequence once as a shared context while preserving stage-specific candidate features and computation. Joint training lets the three stages reinforce one another and enables in-model knowledge distillation from fine-rank to pre-rank. We scale the shared backbone with sparse mixture-of-experts (MoE), which increases capacity with bounded activated computation, and stabilize scaling with $\mu$P-style parameterization. To consolidate objective-specific retrieval channels, we introduce Decision-Conditioned Generative Retrieval (DCGR). DCGR predicts a decision prefix describing the upcoming interaction and generates items conditioned on it, allowing business objectives to steer a single generative process. Finally, Sequence-Native Training (SNT) organizes training around each user's lifelong behavior sequence and amortizes its encoding across exposures. Deployed across all three stages of a large-scale industrial recommendation system, OneTrans-V2 improves gross merchandise value (GMV) by 9.74% and, with a co-designed serving stack, delivers $3.2\times$ the throughput of the cascade it replaces under the same hardware budget.