Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accelerating Generative Recommendation via Simple Categorical User Sequence Compression

Qijiong Liu, Lu Fan, Zhongzhou Liu, Xiaoyu Dong, Yuankai Luo, Guoyuan An, Nuo Chen, Wei Guo, Yong Liu, Xiao-Ming Wu

TL;DR

This work tackles the efficiency bottleneck of generative recommenders that model ultra-long user histories by introducing CAUSE, a simple categorical user sequence compression method. CAUSE partitions histories into a long-term component and a recent sequence, then compresses the history into $V$ history tokens by grouping items into item-category buckets and aggregating bucket representations, yielding input tokens that preserve past interests with far lower compute. The approach achieves up to $6\times$ reductions in computational cost and up to $39.14\%$ accuracy gains at comparable cost across two large datasets and backbone models, and it is model-agnostic, compatible with backbones like HSTU and GenRank. The authors also provide ablations and an open-source implementation, underscoring the method's practicality for real-world deployment of efficient generative recommendations.

Abstract

Although generative recommenders demonstrate improved performance with longer sequences, their real-time deployment is hindered by substantial computational costs. To address this challenge, we propose a simple yet effective method for compressing long-term user histories by leveraging inherent item categorical features, thereby preserving user interests while enhancing efficiency. Experiments on two large-scale datasets demonstrate that, compared to the influential HSTU model, our approach achieves up to a 6x reduction in computational cost and up to 39% higher accuracy at comparable cost (i.e., similar sequence length).

Accelerating Generative Recommendation via Simple Categorical User Sequence Compression

TL;DR

This work tackles the efficiency bottleneck of generative recommenders that model ultra-long user histories by introducing CAUSE, a simple categorical user sequence compression method. CAUSE partitions histories into a long-term component and a recent sequence, then compresses the history into history tokens by grouping items into item-category buckets and aggregating bucket representations, yielding input tokens that preserve past interests with far lower compute. The approach achieves up to reductions in computational cost and up to accuracy gains at comparable cost across two large datasets and backbone models, and it is model-agnostic, compatible with backbones like HSTU and GenRank. The authors also provide ablations and an open-source implementation, underscoring the method's practicality for real-world deployment of efficient generative recommendations.

Abstract

Although generative recommenders demonstrate improved performance with longer sequences, their real-time deployment is hindered by substantial computational costs. To address this challenge, we propose a simple yet effective method for compressing long-term user histories by leveraging inherent item categorical features, thereby preserving user interests while enhancing efficiency. Experiments on two large-scale datasets demonstrate that, compared to the influential HSTU model, our approach achieves up to a 6x reduction in computational cost and up to 39% higher accuracy at comparable cost (i.e., similar sequence length).
Paper Structure (10 sections, 8 equations, 2 figures, 4 tables)

This paper contains 10 sections, 8 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Preliminaries and Related Work
  3. Generative Recommendation
  4. Efficiency Challenge
  5. Categorical User Sequence Compression
  6. Experiments
  7. Experimental Setup
  8. Main Results
  9. Ablation Studies
  10. Conclusion

Figures (2)

  • Figure 1: Training efficiency and recommendation performance of HSTU compared with our proposed CAUSE. Scaling trends are highlighted with fitted lines for both methods.
  • Figure 2: Illustration of the proposed user sequence compression framework. The observed user sequence $\mathbf{s}_u^{\leq t}$ is divided into a long-term history $\mathbf{h}_u$ (in red background) and a recent sequence $\mathbf{r}_u^{\leq t}$ (in blue background). The history is substantially compressed based on item categorical features, while the recent sequence, which reflects the user’s current interests, is used to train both next-item prediction and action prediction tasks.