Reproduce “RECAP: Feedback-Driven Streaming Semantic User Profiles for Short-Video 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
Ziyi Zhao, Xiaoyou Zhou, Xiao Lv, Yangyang Li, Chubo He, Zhao Liu, Jiayao Shen, Yuqi Liu and 8 more
RECAP (Feedback-Driven Streaming Semantic User Profiles for Short-Video Recommendation) introduces a closed-loop framework for optimizing continually updated, structured semantic user profiles to enhance short-video recommendation performance. It achieved a 0.139% relative improvement in average application usage time during an online A/B test on Kuaishou's platform.
Key insights
- Decoupling LLM semantic abstraction from a deterministic state machine is crucial for maintaining reliable and consistent structured profile states over time, effectively preventing errors common in purely LLM-managed states.
- Constructing high-confidence, semantically filtered feedback with an LLM judge is essential for effective policy optimization, leading to profiles that are more grounded and generate deeper semantic understanding than those optimized with raw, noisy feedback.
- Reward formulation combining prediction accuracy (BCE) and semantic separation (cosine-margin) provides a more robust signal for profile optimization, driving better alignment between user interests and item content.
- Optimizing with cleaned feedback encourages LLMs to generate more 'motivation-focused' and abstract user interest descriptions, moving beyond superficial content features seen with raw feedback.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.