Reproduce “Towards End-to-End Alignment of User Satisfaction via Questionnaire in 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
Na Li, Jiaqi Yu, Minzhi Xie, Tiantian He, Xiaoxiao Xu, Zixiu Wang, Lantao Hu, Yongqi Liu and 3 more
Kuaishou Technology researchers developed EASQ, a framework that leverages sparse, high-quality questionnaire feedback to align video recommendation models with true user satisfaction in an online learning environment. Online A/B tests demonstrated consistent improvements in user retention, engagement metrics, and positive explicit feedback.
Key insights
- Parameter-efficient fine-tuning (LoRA) at early layers provides a stable and effective mechanism for injecting sparse, high-quality satisfaction signals into the model without overwhelming dense behavioral data.
- Decoupling main ranking and satisfaction alignment tasks with specialized expert networks allows for robust learning from heterogeneous feedback types and improved generalization.
- Adapting DPO for online settings by dynamically updating the reference model enables continuous, real-time alignment with user satisfaction, addressing the static limitations of traditional DPO in dynamic environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.