Reproduce “STAN: Stage-Adaptive Network for Multi-Task Recommendation by Learning User Lifecycle-Based Representation” 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
Wanda Li, Wenhao Zheng, Xuanji Xiao, Suhang Wang
The Stage-Adaptive Network (STAN) models dynamic user lifecycle stages to adapt multi-task recommendation systems, dynamically weighting task objectives based on learned user preferences. This approach led to online improvements of 3.94% in CTR, 3.05% in staytime, and 0.88% in CVR when deployed in an e-commerce setting.
Key insights
- User preferences for recommendation tasks are dynamic and evolve across distinct lifecycle stages, requiring an adaptive multi-task optimization approach rather than static weighting.
- Modeling user preferences as Beta distributions provides a more stable and robust representation of latent user stages, especially when user action histories are limited.
- Incorporating dynamically learned user stage information effectively mitigates negative transfer in multi-task learning by prioritizing tasks relevant to a user's current engagement, improving overall prediction accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.