Reproduce “USD: A User-Intent-Driven Sampling and Dual-Debiasing Framework for Large-Scale Homepage Recommendations” 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
Jiaqi Zheng, Cheng Guo, Yi Cao, Chaoqun Hou, Tong Liu, Bo Zheng
A framework named USD was developed by Alibaba researchers to improve large-scale homepage recommendations by addressing both exposure bias and click bias through user-intent-driven sampling and dual-debiasing. This method achieved online User Click-Through Rate increases of up to 35.4% on Taobao's marketing blocks.
Key insights
- User intent, rather than just click/non-click behavior, is critical for accurately distinguishing genuine user interest from noise introduced by biases in homepage recommendations.
- Capturing user intent for marketing portals through multi-channel engagements (beyond just block clicks) significantly improves the quality of negative samples by filtering out inattentive exposures.
- A dual-debiasing approach that specifically addresses both click bias (portal-driven clicks) and exposure bias (critical negative samples) through intent-aware weighting leads to improved model accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.