Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time

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

TL;DR

The paper tackles duration bias in video watch-time signals by introducing counterfactual watch time (CWT) and a corresponding Counterfactual Watch Model (CWM). By modeling watching as a utility-maximization process with a cost, CWM derives a cost-based transform linking CWT to user interest and optimizes a counterfactual likelihood that accounts for truncated observations. Empirical results on three real datasets and online A/B tests show CWM improves watch-time prediction and relevance ranking while effectively debiasing for video duration. The approach offers a principled, causal framing for extracting true user interest from biased implicit feedback, with practical impact for scalable video recommendation systems.

Abstract

In video recommendation, an ongoing effort is to satisfy users' personalized information needs by leveraging their logged watch time. However, watch time prediction suffers from duration bias, hindering its ability to reflect users' interests accurately. Existing label-correction approaches attempt to uncover user interests through grouping and normalizing observed watch time according to video duration. Although effective to some extent, we found that these approaches regard completely played records (i.e., a user watches the entire video) as equally high interest, which deviates from what we observed on real datasets: users have varied explicit feedback proportion when completely playing videos. In this paper, we introduce the counterfactual watch time(CWT), the potential watch time a user would spend on the video if its duration is sufficiently long. Analysis shows that the duration bias is caused by the truncation of CWT due to the video duration limitation, which usually occurs on those completely played records. Besides, a Counterfactual Watch Model (CWM) is proposed, revealing that CWT equals the time users get the maximum benefit from video recommender systems. Moreover, a cost-based transform function is defined to transform the CWT into the estimation of user interest, and the model can be learned by optimizing a counterfactual likelihood function defined over observed user watch times. Extensive experiments on three real video recommendation datasets and online A/B testing demonstrated that CWM effectively enhanced video recommendation accuracy and counteracted the duration bias.

Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time

TL;DR

The paper tackles duration bias in video watch-time signals by introducing counterfactual watch time (CWT) and a corresponding Counterfactual Watch Model (CWM). By modeling watching as a utility-maximization process with a cost, CWM derives a cost-based transform linking CWT to user interest and optimizes a counterfactual likelihood that accounts for truncated observations. Empirical results on three real datasets and online A/B tests show CWM improves watch-time prediction and relevance ranking while effectively debiasing for video duration. The approach offers a principled, causal framing for extracting true user interest from biased implicit feedback, with practical impact for scalable video recommendation systems.

Abstract

In video recommendation, an ongoing effort is to satisfy users' personalized information needs by leveraging their logged watch time. However, watch time prediction suffers from duration bias, hindering its ability to reflect users' interests accurately. Existing label-correction approaches attempt to uncover user interests through grouping and normalizing observed watch time according to video duration. Although effective to some extent, we found that these approaches regard completely played records (i.e., a user watches the entire video) as equally high interest, which deviates from what we observed on real datasets: users have varied explicit feedback proportion when completely playing videos. In this paper, we introduce the counterfactual watch time(CWT), the potential watch time a user would spend on the video if its duration is sufficiently long. Analysis shows that the duration bias is caused by the truncation of CWT due to the video duration limitation, which usually occurs on those completely played records. Besides, a Counterfactual Watch Model (CWM) is proposed, revealing that CWT equals the time users get the maximum benefit from video recommender systems. Moreover, a cost-based transform function is defined to transform the CWT into the estimation of user interest, and the model can be learned by optimizing a counterfactual likelihood function defined over observed user watch times. Extensive experiments on three real video recommendation datasets and online A/B testing demonstrated that CWM effectively enhanced video recommendation accuracy and counteracted the duration bias.
Paper Structure (33 sections, 1 theorem, 13 equations, 12 figures, 5 tables)

This paper contains 33 sections, 1 theorem, 13 equations, 12 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Counterfactual watch time
  4. Definition of counterfactual watch time
  5. The existence of counterfactual watch time
  6. Evidence 1: repeated playing
  7. Evidence 2: bimodal distribution
  8. Explanation from counterfactual watch time
  9. An economic view of user watching
  10. Limitation of existing methods
  11. Our approach
  12. Cost-based transform function
  13. Counterfactual likelihood function
  14. Formulation of the counterfactual likelihood function
  15. Parameterize and optimize the likelihood function
  16. ...and 18 more sections

Key Result

Theorem 1

For $\forall~~\mathcal{W} \subseteq \mathbb{R}^+, g\in\mathcal{G}$, given $g: \mathcal{R} \to \mathcal{W}$, we have $\nexists~~ g^{-1}: \mathcal{W} \to \mathcal{R}$, where $\mathcal{W}$ is the set of all observed watch time values, $\mathcal{G}$ is the function space, $\mathcal{R}$ is the set of a

Figures (12)

  • Figure 1: Users' explicit feedback proportion in completely played records grouped by video duration of (a) KuaiRand dataset (b) WeChat dataset.
  • Figure 2: Comparison between counterfactual watch time and observed watch time. Users A and B have the same observed watch time but different counterfactual watch times.
  • Figure 3: The repeated play proportion and average repeated play ratio in different duration bins of (a) KuaiRand dataset (b) WeChat dataset.
  • Figure 4: The bimodal distribution of watch time on (a) KuaiRand dataset (b) WeChat dataset. The video duration is 30s.
  • Figure 5: Employing CWT to explain (a) repeated playing (b) bimodal distribution.
  • ...and 7 more figures

Theorems & Definitions (2)

  • Definition 1: counterfactual watch time
  • Theorem 1: observed watch time is not the indicator of interest