Reproduce “STEM: Unleashing the Power of Embeddings for Multi-task Recommendation” 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
Liangcai Su, Junwei Pan, Ximei Wang, Xi Xiao, Shijie Quan, Xihua Chen, Jie Jiang
This research introduces the Shared and Task-specific EMbeddings (STEM) paradigm and the STEM-Net model, which improve multi-task recommendation by explicitly learning both shared and task-specific embeddings for users and items. The approach addresses negative transfer observed in existing models on 'comparable samples' and achieved up to 7.1% Gross Merchandise Value (GMV) lift in real-world deployment on Tencent's online advertising platform.
Key insights
- Negative transfer in multi-task learning (MTL) is most pronounced and detrimental on 'comparable samples' where users exhibit balanced positive feedback across tasks, contrary to the general assumption of universal negative transfer.
- The traditional shared-embedding paradigm fundamentally limits MTL models from capturing diverse or contradictory user preferences for items across different tasks.
- Explicitly learning task-specific embeddings alongside shared embeddings is critical for overcoming negative transfer and effectively modeling nuanced user behaviors in multi-task recommendation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.