Reproduce “Rethinking Multi-Interest Learning for Candidate Matching in Recommender Systems” 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
Yueqi Xie, Jingqi Gao, Peilin Zhou, Qichen Ye, Yining Hua, Jaeboum Kim, Fangzhao Wu, Sunghun Kim
The paper 'Rethinking Multi-Interest Learning for Candidate Matching in Recommender Systems' introduces REMI, a general training framework addressing two key limitations in multi-interest learning: ineffective negative sampling and routing collapse. REMI significantly improves the performance of existing multi-interest models across various metrics on large-scale datasets, achieving up to 19.7% relative improvement in Recall@20 on Amazon Books.
Key insights
- Fundamental improvements in training dynamics, particularly negative sampling and routing mechanisms, can yield substantial performance gains for multi-interest learning models.
- Uniform negative sampling is suboptimal for multi-interest learning due to the 'selected-interest-focused training,' which generates many 'easy negatives' that provide limited discriminative signal.
- Multi-interest routing mechanisms can suffer from 'routing collapse,' where learned interests over-focus on single dominant historical items instead of aggregating information from diverse sets of relevant items.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.