Conditional Quantile Estimation for Uncertain Watch Time in Short-Video Recommendation

TL;DR

The paper tackles uncertainty in watch-time prediction for short-video recommendations by modeling the full conditional distribution $P(W|oldsymbol{x})$ via Conditional Quantile Estimation (CQE) using quantile regression. CQE outputs multiple quantiles ${t_{\tau_i}}$ through a monotone neural network and is trained with the pinball loss across quantiles, enabling three inference strategies: Conservative Estimation, Dynamic Quantile Combination, and Conditional Expectation. Online A/B tests on a large-scale platform show gains in active days, engagement time, and video views across CQE strategies, while offline experiments on public datasets corroborate improvements in watch-time prediction and user-interest modeling across various backbones. The results demonstrate that distribution-aware CQE provides scalable, flexible personalization for short-video recommender systems, with code released for reproducibility on GitHub.

Abstract

Accurately predicting watch time is crucial for optimizing recommendations and user experience in short video platforms. However, existing methods that estimate a single average watch time often fail to capture the inherent uncertainty in user engagement patterns. In this paper, we propose Conditional Quantile Estimation (CQE) to model the entire conditional distribution of watch time. Using quantile regression, CQE characterizes the complex watch-time distribution for each user-video pair, providing a flexible and comprehensive approach to understanding user behavior. We further design multiple strategies to combine the quantile estimates, adapting to different recommendation scenarios and user preferences. Extensive offline experiments and online A/B tests demonstrate the superiority of CQE in watch-time prediction and user engagement modeling. Specifically, deploying CQE online on a large-scale platform with hundreds of millions of daily active users has led to substantial gains in key evaluation metrics, including active days, engagement time, and video views. These results highlight the practical impact of our proposed approach in enhancing the user experience and overall performance of the short video recommendation system. The code will be released https://github.com/justopit/CQE.

Figures (5)

• Figure 1: Illustration of the conditional watch time distribution and some personalized recommendation strategies. (a) Conditional watch time distributions predicted by our CQE model for different user-video pairs, showcasing the heterogeneity and complexity of user engagement patterns. (b) Conservative estimation strategy prioritizing user satisfaction by selecting videos with higher lower quantiles when the expected watch times are similar. (c) Dynamic quantile combination strategy adapting to user churn risk or video novelty, employing lower quantiles for high-churn-risk users or unfamiliar videos, and higher quantiles for low-churn-risk users or familiar videos. (d) Expectation estimation strategy providing a global optimization perspective by considering the entire watch time distribution. We normalize watch time by limiting it to a maximum of 300 seconds and scaling it as min(watch time, 300) / 300.
• Figure 2: Illustration of the proposed Conditional Quantile Estimation (CQE) model and its training loss. Left: The CQE model architecture, which takes user, video, and context features as input and outputs multiple quantile estimates of watch time. Right: The pinball loss function used for training the model, showing its asymmetric nature for different quantile levels ($\tau$), allowing the model to learn robust quantile estimates across the entire distribution.
• Figure 3: The diversity performances of CQEDQC compared with baseline online.
• Figure 4: The performances of CQECDE with various number of quantiles in CIKM16 Dataset for predicting watch time.
• Figure 5: The performances of CQECDE with various number of quantiles in KuaiRand Dataset for predicting user interest.