Reproduce “Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time” 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

Haiyuan Zhao, Guohao Cai, Jieming Zhu, Zhenhua Dong, Jun Xu, Ji-Rong Wen

The Counterfactual Watch Model (CWM) introduces Counterfactual Watch Time (CWT) to mitigate duration bias in video recommendation by modeling user interest as an untruncated potential watch time. This approach significantly improves watch time prediction and relevance ranking accuracy across datasets and boosts online user engagement metrics.

Key insights

  • Duration bias is reinterpreted as the truncation of a user's true desired watch time (CWT) by the actual video duration.
  • The concept of CWT provides a unified explanation for phenomena such as users repeatedly playing short videos (when CWT exceeds video length) and the observed bimodal distribution of watch times.
  • A formal theorem establishes that uniquely inferring true user interest from observed watch time alone is impossible for completely played records, highlighting the necessity of explicitly modeling truncation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.