Reproduce “PEAR: Personalized Re-ranking with Contextualized Transformer for 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
Yi Li, Jieming Zhu, Weiwen Liu, Liangcai Su, Guohao Cai, Qi Zhang, Ruiming Tang, Xi Xiao and 1 more
The PEAR framework introduces a personalized re-ranking model for recommender systems that integrates a user's historical clicked item list through a contextualized transformer architecture and a multi-task learning approach. This method enhances the refinement of initial recommendation lists, achieving up to a 1.71% improvement in gAUC@30 on a news dataset and 1.07% on a micro-video dataset compared to competitive baselines.
Key insights
- Explicitly leveraging a user's historical clicked item list significantly enhances personalized re-ranking by providing richer contextual information beyond the immediate candidate list.
- The merged cross-attention mechanism efficiently captures complex item-item interactions both within the target list and across the target and historical behavior lists.
- Incorporating a list-level auxiliary task during training provides a beneficial global supervision signal, contributing to the model's overall effectiveness in optimizing list quality.
- Performance generally improves with longer historical item sequences, indicating the importance of comprehensive sequential signals for effective personalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.