HiGR: Efficient Generative Slate Recommendation
Analysis
This paper introduces HiGR, a novel framework for slate recommendation that addresses limitations in existing autoregressive models. It focuses on improving efficiency and recommendation quality by integrating hierarchical planning and preference alignment. The key contributions are a structured item tokenization method, a two-stage generation process (list-level planning and item-level decoding), and a listwise preference alignment objective. The results show significant improvements in both offline and online evaluations, highlighting the practical impact of the proposed approach.
Key Takeaways
- •Proposes HiGR, a novel framework for slate recommendation.
- •Integrates hierarchical planning and listwise preference alignment.
- •Achieves significant improvements in both offline and online evaluations.
- •Offers a 5x inference speedup compared to state-of-the-art methods.
“HiGR delivers consistent improvements in both offline evaluations and online deployment. Specifically, it outperforms state-of-the-art methods by over 10% in offline recommendation quality with a 5x inference speedup, while further achieving a 1.22% and 1.73% increase in Average Watch Time and Average Video Views in online A/B tests.”