Reproduce “Unified Embedding Based Personalized Retrieval in Etsy Search” 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
Rishikesh Jha, Siddharth Subramaniyam, Ethan Benjamin, Thrivikrama Taula
Etsy Inc. developed a Unified Embedding Based Personalized Product Retrieval (UEPPR) system that combines graph, transformer, and term-based embeddings within a two-tower neural architecture. This system, deployed at industry scale, improved Etsy's site-wide conversion rate by 2.63% and organic search purchase rate by 5.58%.
Key insights
- Combining diverse embedding types (graph, transformer, token, location) within a unified model provides comprehensive semantic understanding and personalization, with graph embeddings yielding the most substantial recall improvement.
- Dynamic, weighted hard negative mining is critical for robust training of dense retrieval models, delivering an 11% relative gain in Recall@100 compared to basic negative sampling.
- Integrating query-independent product quality directly into the retrieval stage via ANN-Based Product Boosting significantly optimizes the initial candidate set, especially for head queries, and learning these weights through black-box optimization prevents overfitting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.