Reproduce “Mosaic: A Fleet of User Embedding Specialists for Recommendation at Meta” 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
John Zhiyuan Zheng, Xian Sun, Xiangyang Mou, Yujunrong Ma, Christina You, Michael Jiayuan He, Hrishikesh Paranjape, Aakarsha Agarwal and 1 more
Meta's Mosaic system employs a fleet of specialized user embedding models to enhance recommendation quality by capturing diverse user behaviors more effectively. This approach delivers consistent improvements in offline Normalized Entropy and online engagement metrics across multiple product surfaces.
Key insights
- A fleet of specialized user embedding models, each with tailored inductive biases, is more effective at capturing heterogeneous user behaviors than a single generalist model.
- Decoupling user representation learning into dedicated upstream models enhances reusability, scalability, and cross-surface signal transfer across various downstream recommendation consumers.
- Specific strategies like redundancy loss and multi-task relation mining are critical for ensuring that each new specialist contributes unique, non-redundant information to a growing fleet.
- Accelerated iteration velocity through logging-free offline evaluation methods (like CoEval) is crucial for efficient development and deployment of large-scale user embedding systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.