Reproduce “RankGR: Rank-Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kairui Fu, Changfa Wu, Kun Yuan, Binbin Cao, Dunxian Huang, Yuliang Yan, Junjun Zheng, Jianning Zhang and 3 more
A collaborative effort by Zhejiang University and Alibaba Group, RankGR enhances generative retrieval by integrating listwise direct preference optimization for hierarchical user preferences and a refined scoring phase for deep candidate interaction. The system achieved state-of-the-art offline performance and, in online A/B tests on Taobao, increased Item Page Views by 1.08% and Transaction Count by 0.57%.
Key insights
- Explicitly modeling hierarchical user preferences through Listwise Direct Preference Optimization (LDPO) is crucial, leading to an 18.28% improvement in Click HR@20 compared to not using it.
- The Refined Scoring Phase (RSP) significantly enhances precision for top-K recommendations by enabling deep, candidate-centric interactions, evidenced by a 10.44% drop in Click HR@20 without it.
- An optimal balance between the traditional next-token prediction and the novel preference alignment objective is essential for maximizing the model's performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.