SWaT: Statistical Modeling of Video Watch Time through User Behavior Analysis

Abstract: The significance of estimating video watch time has been highlighted by the rising importance of (short) video recommendation, which has become a core product of mainstream social media platforms. Modeling video watch time, however, has been challenged by the complexity of user-video interaction, such as different user behavior modes in watching the recommended videos and varying watching probability over the video progress bar. Despite the importance and challenges, existing literature on modeling video watch time mostly focuses on relatively black-box mechanical enhancement of the classical regression/classification losses, without factoring in user behavior in a principled manner. In this paper, we for the first time take on a user-centric perspective to model video watch time, from which we propose a white-box statistical framework that directly translates various user behavior assumptions in watching (short) videos into statistical watch time models. These behavior assumptions are portrayed by our domain knowledge on users' behavior modes in video watching. We further employ bucketization to cope with user's non-stationary watching probability over the video progress bar, which additionally helps to respect the constraint of video length and facilitate the practical compatibility between the continuous regression event of watch time and other binary classification events. We test our models extensively on two public datasets, a large-scale offline industrial dataset, and an online A/B test on a short video platform with hundreds of millions of daily-active users. On all experiments, our models perform competitively against strong relevant baselines, demonstrating the efficacy of our user-centric perspective and proposed framework.

• The paper presents a user-centric statistical framework that models video watch time by leveraging distinct user behavior modes.
• It employs a bucketization strategy and both binomial and geometric models to capture non-stationary watching probabilities.
• Experimental validation on public and industrial datasets demonstrates significant improvements in prediction accuracy and user engagement.

Statistical Modeling of Video Watch Time through User Behavior Analysis

The estimation of video watch time is a crucial metric for social media platforms, particularly given the increasing popularity of short video content. This paper takes a user-centric perspective to model video watch time, addressing the complexity of user-video interactions and overcoming the deficiencies of traditional regression and classification methods. The authors propose a white-box statistical framework that translates user behavior assumptions into statistical models, effectively capturing the nuances of different user behavior modes and the non-stationary probabilities over the video horizon.

Major Contributions

The paper's primary contribution is the development of a statistical framework that leverages user behavior assumptions to model video watch time. This approach diverges from conventional methods that largely ignore user behavior, focusing instead on enhancing regression/classification objectives mechanically. The framework introduced in this paper enhances understanding of user behavior by embodying domain knowledge about users' behavior modes in watching videos.

Key aspects of the proposed framework include:

1. User-Centric Perspective: The framework incorporates domain knowledge about user behavior, distinguishing between different behavior modes (e.g., wandering-minded users who jump around the video timeline and focused users who watch sequentially).
2. Bucketization Strategy: This is employed to handle non-stationary watching probabilities over the video horizon, facilitating practical modeling and respecting the constraints of video length.

Methodology

The authors present two main models within their framework: the _-Binom and the _-Geo models.

_-Binom Model

This model assumes users have wandering behavior, where the decision to watch each second within a bucket is made independently with an identical probability. The model utilizes bucketization to split the video horizon into buckets, estimating a separate probability for each bucket. This results in a binomial distribution model for watch time that is straightforward to implement and is compatible with training objectives of other classification events in practical systems.

_-Geo Model

In contrast, the _-Geo model assumes users exhibit sequential watching behavior with a stopping probability for each second. Two variants of this model are proposed:

• The _-Geo model allows for non-stationary watching probabilities, specific to each bucket.
• The vanilla geometric model (_v-geo) simplifies this by assuming a uniform watching probability across the entire horizon.

The expectation of video watch time in these models is derived analytically and efficiently computes closed-form solutions.

Experimental Validation

The authors validate their models through extensive experiments on multiple datasets:

1. Public Datasets (CIKM16 Cup and KuaiRec): The proposed models consistently outperformed state-of-the-art methods such as Weighted Logistic Regression (WLR), Duration-Deconfounded Quantile Regression (D2Q), Ordinal Regression (OR), and Tree-based Progressive Regression Model (TPM) on metrics like Mean Absolute Error (MAE) and XAUC.
2. Large-Scale Industrial Dataset: When evaluated on a dataset from a popular short video platform, the _-Binom and _-Geo models showed a significant performance improvement over the baseline prediction models, based on a large corpus of user interactions.
3. Online A/B Test: In real-world deployment on a platform with hundreds of millions of daily-active users, the proposed methods increased both the total watch time and watch count significantly, corroborating the practical efficacy of their approach.

Implications and Future Directions

The framework proposed in this paper has both theoretical and practical implications. The integration of user behavior into statistical models offers a more nuanced understanding and prediction of user engagement, which is critical for optimizing video recommendation systems. By leveraging domain knowledge and bucketization, the models are both interpretable and compatible with existing industrial systems.

Future research may focus on refining the user behavior assumptions and extending the modeling to more varied types of video content and user interactions. Additionally, the bucketization strategy opens avenues for further studies on optimal bucket selection and adaptive bucketization techniques, potentially improving the model's flexibility and robustness.

Conclusion

This paper presents a comprehensive and user-centric approach to modeling video watch time, addressing the intricate user-video interactions that complicate such predictions. The proposed statistical models are innovative in their integration of domain knowledge and effective in improving prediction accuracy and user engagement metrics. This research provides a solid foundation for future advancements in video recommendation systems and user behavior analysis.