Reproduce “Perceive Your Users in Depth: Learning Universal User Representations from Multiple E-commerce Tasks” on OpenResearch

About this paper

Yabo Ni, Dan Ou, Shichen Liu, Xiang Li, Wenwu Ou, Anxiang Zeng, Luo Si

Alibaba Group researchers developed the Deep User Perception Network (DUPN), a multi-task learning framework designed to learn universal user representations from diverse e-commerce behaviors. Deployed in Taobao's search system, DUPN achieved a 2.23% increase in click-through rate and a 3.17% increase in sales volume, while also demonstrating improved efficiency and representation transferability across tasks.

Key insights

• Explicitly incorporating behavior properties into the deep learning architecture, particularly within the LSTM gates and attention mechanism, significantly enhances the quality and contextual relevance of learned user representations.
• Jointly training a model on multiple e-commerce tasks (e.g., CTR prediction, ranking, price preference) within a multi-task learning framework leads to faster convergence and consistently higher accuracy for each individual task compared to single-task models.
• The learned universal user representations are highly generalizable and transferable, allowing them to be effectively utilized in new, previously unseen personalization tasks with minimal effort, thus accelerating feature development.

Summary by alphaXiv.