Reproduce “Que2Engage: Embedding-based Retrieval for Relevant and Engaging Products at Facebook Marketplace” 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
Yunzhong He, Yuxin Tian, Mengjiao Wang, Feier Chen, Licheng Yu, Maolong Tang, Congcong Chen, Ning Zhang and 2 more
Que2Engage, developed by Meta and UC Merced, addresses the misalignment between semantic relevance and user engagement in e-commerce search by integrating contextual information into the retrieval stage. The system, deployed on Facebook Marketplace, uses a multimodal architecture with a multitask training strategy to optimize for both objectives, resulting in a 4.5% improvement in searcher engagement during online A/B testing.
Key insights
- Integrating contextual features as a distinct modality within a multimodal transformer fusion architecture effectively bridges the relevance-engagement gap in e-commerce search.
- A multitask learning strategy combining a contrastive relevance loss with an engagement loss (using hard negatives of displayed-but-unengaged products) significantly improves user engagement.
- Modality dropout enhances model robustness and performance by preventing over-reliance on any single information source.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.