Reproduce “Bootstrapping Conditional Retrieval for User-to-Item 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
Hongtao Lin, Haoyu Chen, Jaewon Jang, Jiajing Xu
Pinterest researchers developed Conditional Retrieval (CR), a method that extends two-tower recommendation models to incorporate new conditions efficiently without requiring condition-specific training data. This approach, evaluated at Pinterest, enhanced user engagement metrics like Email CTR by +2.86% and drastically cut annual infrastructure costs compared to standard baseline retrieval methods.
Key insights
- Conditional retrieval can be effectively bootstrapped for new conditions by deriving them from existing item-side features, bypassing the need for specific engagement data.
- Deeply integrating conditions into the user tower's learning process enables the model to holistically balance user engagement with condition relevance.
- Learning condition relevance during training leads to significantly lower serving costs due to higher inherent topic matching rates, reducing the need for costly post-filtering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.