Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OneMall: One Model, More Scenarios -- End-to-End Generative Recommender Family at Kuaishou E-Commerce

Kun Zhang, Jingming Zhang, Wei Cheng, Yansong Cheng, Jiaqi Zhang, Hao Lu, Xu Zhang, Haixiang Gan, Jiangxia Cao, Tenglong Wang, Ximing Zhang, Boyang Xia, Kuo Cai, Shiyao Wang, Hongjian Dou, Jinkai Yu, Mingxing Wen, Qiang Luo, Dongxu Liang, Chenyi Lei, Jun Wang, Runan Liu, Zhaojie Liu, Ruiming Tang, Tingting Gao, Shaoguo Liu, Yuqing Ding, Hui Kong, Han Li, Guorui Zhou, Wenwu Ou, Kun Gai

TL;DR

OneMall presents an end-to-end generative recommender for e-commerce that unifies three item scenarios (product-card, short-video, live-streaming) using a semantic tokenizer, a pure Transformer backbone with long-sequence compression and sparse MoE, and an online RL framework to couple retrieval with ranking. It introduces Res-Kmeans and FSQ-based semantic IDs to capture real-world semantics and business relations, and uses a decoder-only architecture with Query-Former and Cross-Attention to fuse user histories. Across offline, simulation, and online experiments, OneMall achieves consistent gains in GMV and orders across all scenarios and is deployed to hundreds of millions of users, demonstrating the practicality of adapting LLM-style end-to-end techniques to revenue-oriented e-commerce. The work also provides extensive ablations on tokenizer design, sequence compression, and RL strategies, offering actionable insights for scaling and unifying retrieval and ranking in industrial RecSys settings.

Abstract

In the wave of generative recommendation, we present OneMall, an end-to-end generative recommendation framework tailored for e-commerce services at Kuaishou. Our OneMall systematically unifies the e-commerce's multiple item distribution scenarios, such as Product-card, short-video and live-streaming. Specifically, it comprises three key components, aligning the entire model training pipeline to the LLM's pre-training/post-training: (1) E-commerce Semantic Tokenizer: we provide a tokenizer solution that captures both real-world semantics and business-specific item relations across different scenarios; (2) Transformer-based Architecture: we largely utilize Transformer as our model backbone, e.g., employing Query-Former for long sequence compression, Cross-Attention for multi-behavior sequence fusion, and Sparse MoE for scalable auto-regressive generation; (3) Reinforcement Learning Pipeline: we further connect retrieval and ranking models via RL, enabling the ranking model to serve as a reward signal for end-to-end policy retrieval model optimization. Extensive experiments demonstrate that OneMall achieves consistent improvements across all e-commerce scenarios: +13.01\% GMV in product-card, +15.32\% Orders in Short-Video, and +2.78\% Orders in Live-Streaming. OneMall has been deployed, serving over 400 million daily active users at Kuaishou.

OneMall: One Model, More Scenarios -- End-to-End Generative Recommender Family at Kuaishou E-Commerce

TL;DR

OneMall presents an end-to-end generative recommender for e-commerce that unifies three item scenarios (product-card, short-video, live-streaming) using a semantic tokenizer, a pure Transformer backbone with long-sequence compression and sparse MoE, and an online RL framework to couple retrieval with ranking. It introduces Res-Kmeans and FSQ-based semantic IDs to capture real-world semantics and business relations, and uses a decoder-only architecture with Query-Former and Cross-Attention to fuse user histories. Across offline, simulation, and online experiments, OneMall achieves consistent gains in GMV and orders across all scenarios and is deployed to hundreds of millions of users, demonstrating the practicality of adapting LLM-style end-to-end techniques to revenue-oriented e-commerce. The work also provides extensive ablations on tokenizer design, sequence compression, and RL strategies, offering actionable insights for scaling and unifying retrieval and ranking in industrial RecSys settings.

Abstract

In the wave of generative recommendation, we present OneMall, an end-to-end generative recommendation framework tailored for e-commerce services at Kuaishou. Our OneMall systematically unifies the e-commerce's multiple item distribution scenarios, such as Product-card, short-video and live-streaming. Specifically, it comprises three key components, aligning the entire model training pipeline to the LLM's pre-training/post-training: (1) E-commerce Semantic Tokenizer: we provide a tokenizer solution that captures both real-world semantics and business-specific item relations across different scenarios; (2) Transformer-based Architecture: we largely utilize Transformer as our model backbone, e.g., employing Query-Former for long sequence compression, Cross-Attention for multi-behavior sequence fusion, and Sparse MoE for scalable auto-regressive generation; (3) Reinforcement Learning Pipeline: we further connect retrieval and ranking models via RL, enabling the ranking model to serve as a reward signal for end-to-end policy retrieval model optimization. Extensive experiments demonstrate that OneMall achieves consistent improvements across all e-commerce scenarios: +13.01\% GMV in product-card, +15.32\% Orders in Short-Video, and +2.78\% Orders in Live-Streaming. OneMall has been deployed, serving over 400 million daily active users at Kuaishou.
Paper Structure (18 sections, 1 equation, 7 figures, 6 tables)

This paper contains 18 sections, 1 equation, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Feature, Retrieval and Ranking Pipeline
  4. Res-Kmeans Semantic Tokenizer
  5. OneMall Workflow
  6. Semantic Tokenizer for E-commerce
  7. Business Item2Item Data Filter
  8. Semantic Compression Backbone
  9. Semantic ID Generation Pipeline
  10. Decoder-Only Model Architecture
  11. Reinforcement Learning Policy
  12. Experiments
  13. Offline Scaling Experiments
  14. Simulation Experiments
  15. Online Experiments
  16. ...and 3 more sections

Figures (7)

  • Figure 1: Three major shopping item categories at Kuaishou, where the product-card represents one specific product, the short-video typically connects to only one product-card, live-streaming can sell multiple product-cards simultaneously.
  • Figure 2: The model structure in industrial RecSys chain: (a) two-tower retrieval model; (b) more feature ranking model; (c) RecSys chain, first pass the user tower of retrieval model, and then feed to ranking model to get a fusion score.
  • Figure 3: (a) LLM fine-tuning process in Product-Card/Short-Video Item2Item dataset; (c) Live-Streaming embedding generation with selling product Semantic IDs; (b) ResKmeansFSQ Tokenizer, different scenarios utilize different embedding candidates.
  • Figure 4: The Transformer backbone of OneMall.
  • Figure 5: The RL workflow of our OneMall.
  • ...and 2 more figures