Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Large-scale Generative Ranking

Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, Yuting Jia, Leilei Ma, Yinqi Zhang, Taoyu Zhu, Liujie Zhang, Lei Chen, Weihang Chen, Min Zhu, Ruiwen Xu, Lei Zhang

TL;DR

The paper addresses the challenge of applying generative ranking in large-scale industrial recommender systems. It argues that the generative architecture is the main driver of performance rather than training paradigms, and introduces GenRank, which uses an action-oriented organization and linear position/time biases to improve efficiency. Extensive offline analyses and large-scale online A/B tests in Xiaohongshu's Explore Feed show GenRank achieves significant gains with comparable resource usage, including improved P99 latency and better performance on cold-start items thanks to content embeddings. The work demonstrates that generative ranking can be both effective and scalable in practice, suggesting a direction toward unifying ranking and pre-ranking with efficient architectures.

Abstract

Generative recommendation has recently emerged as a promising paradigm in information retrieval. However, generative ranking systems are still understudied, particularly with respect to their effectiveness and feasibility in large-scale industrial settings. This paper investigates this topic at the ranking stage of Xiaohongshu's Explore Feed, a recommender system that serves hundreds of millions of users. Specifically, we first examine how generative ranking outperforms current industrial recommenders. Through theoretical and empirical analyses, we find that the primary improvement in effectiveness stems from the generative architecture, rather than the training paradigm. To facilitate efficient deployment of generative ranking, we introduce GenRank, a novel generative architecture for ranking. We validate the effectiveness and efficiency of our solution through online A/B experiments. The results show that GenRank achieves significant improvements in user satisfaction with nearly equivalent computational resources compared to the existing production system.

Towards Large-scale Generative Ranking

TL;DR

The paper addresses the challenge of applying generative ranking in large-scale industrial recommender systems. It argues that the generative architecture is the main driver of performance rather than training paradigms, and introduces GenRank, which uses an action-oriented organization and linear position/time biases to improve efficiency. Extensive offline analyses and large-scale online A/B tests in Xiaohongshu's Explore Feed show GenRank achieves significant gains with comparable resource usage, including improved P99 latency and better performance on cold-start items thanks to content embeddings. The work demonstrates that generative ranking can be both effective and scalable in practice, suggesting a direction toward unifying ranking and pre-ranking with efficient architectures.

Abstract

Generative recommendation has recently emerged as a promising paradigm in information retrieval. However, generative ranking systems are still understudied, particularly with respect to their effectiveness and feasibility in large-scale industrial settings. This paper investigates this topic at the ranking stage of Xiaohongshu's Explore Feed, a recommender system that serves hundreds of millions of users. Specifically, we first examine how generative ranking outperforms current industrial recommenders. Through theoretical and empirical analyses, we find that the primary improvement in effectiveness stems from the generative architecture, rather than the training paradigm. To facilitate efficient deployment of generative ranking, we introduce GenRank, a novel generative architecture for ranking. We validate the effectiveness and efficiency of our solution through online A/B experiments. The results show that GenRank achieves significant improvements in user satisfaction with nearly equivalent computational resources compared to the existing production system.
Paper Structure (13 sections, 1 equation, 3 figures, 2 tables)

This paper contains 13 sections, 1 equation, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem Setup
  3. Source of Effectiveness in Generative Recommendation
  4. Key Mechanisms in Generative Paradigm
  5. Module Performance Comparison across Paradigms
  6. Efficient Generative Ranking in Industrial Scenarios
  7. Item-Action Organization
  8. Position & Time Biases
  9. Related Work
  10. Generative Recommendation
  11. Scaling Law in Recommendation System
  12. Online Experiments
  13. Conclusion

Figures (3)

  • Figure 1: Left: A screenshot of Xiaohongshu's Explore Feed product. Right: Illustration of cascade pipeline for industrial recommender systems, where each stage needs to process a large number of items.
  • Figure 2: Model architecture of GenRank. Compared to existing approaches, e.g., HSTU zhai2024actions, which adopt an item-oriented organization, our solution adopts an action-oriented organization.
  • Figure 3: Illustration of input representation and candidate mask. (a) The input representation of GenRank contains five kinds of embeddings. (b) Candidate Mask: We depict the mask structure for a user's sequence with one historical behavior and two candidate items.