Reproduce “Deconfounding Duration Bias in Watch-time Prediction for 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
Ruohan Zhan, Changhua Pei, Qiang Su, Jianfeng Wen, Xueliang Wang, Guanyu Mu, Dong Zheng, Peng Jiang
Researchers introduced a causal framework to deconfound duration bias in watch-time prediction for video recommendations, which systematically favors longer videos. Their D2Q framework, particularly the Res-D2Q variant, boosted total WatchTime by +0.746% in live A/B tests on the Kuaishou App while maintaining other engagement metrics.
Key insights
- Video duration acts as a confounder, simultaneously affecting user watch-time and the likelihood of a video being exposed by the recommender system.
- Applying backdoor adjustment by stratifying data based on duration quantiles is an effective strategy to mitigate the spurious duration bias while preserving the natural effect of duration on watch-time.
- Predicting watch-time quantiles instead of raw watch-time within duration groups enables a scalable solution that maintains deconfounding benefits with shared model parameters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.