Reproduce “USER: A Unified Information Search and Recommendation Model based on Integrated Behavior Sequence” 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
Jing Yao, Zhicheng Dou, Ruobing Xie, Yanxiong Lu, Zhiping Wang, Ji-Rong Wen
The USER model integrates user search and recommendation behaviors into a single sequence, leveraging a multi-level Transformer architecture to create a unified framework for both tasks. This approach achieves up to 3.49% MAP improvement for personalized search and 1.02% MAP improvement for recommendation over strong baselines on a real-world dataset, particularly in cold-start and mixed-behavior scenarios.
Key insights
- Integrating heterogeneous user behaviors from both search and recommendation tasks into a single, comprehensive sequence provides a more complete and accurate user interest profile than relying on isolated task data.
- Explicitly modeling user intent at both short-term (session-level consistency) and long-term (stable interests) using a multi-level Transformer architecture is crucial for clarifying evolving information needs.
- A unified training approach, combining pre-training on mixed data with task-specific fine-tuning, effectively leverages increased data volume and mitigates data sparsity challenges, especially benefiting cold-start scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.