Reproduce “Tweedie Regression for Video Recommendation System” 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
Yan Zheng, Qiang Chen, Chenglei Niu
Researchers at Tubi implemented Tweedie regression for video recommendation, directly optimizing for user watch time and revenue rather than traditional click-through rates. Their online A/B tests showed a +0.4% increase in revenue and a +0.15% lift in total viewing time.
Key insights
- Watch time data in video-on-demand (VOD) platforms often exhibits a 'mass at zero' and a positively skewed continuous distribution for positive values, closely resembling a Tweedie distribution.
- Tweedie regression, specifically with a power parameter of p=1.5 (representing a compound Poisson-gamma process), provides a theoretically sound and empirically superior method for directly optimizing user watch time compared to traditional classification or ordinary regression approaches.
- Directly optimizing for metrics like watch time and revenue can lead to a nuanced trade-off, potentially increasing deep engagement and revenue from a subset of users while slightly decreasing the overall conversion rate (users watching at least 5 minutes).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.