DeLoad: Demand-Driven Short-Video Preloading with Scalable Watch-Time Estimation

Tong Liu, Zhiwei Fan, Guanyan Peng, Haodan Zhang, Yucheng Zhang, Zhen Wang, Pengjin Xie, Liang Liu

Introduction

In the rapidly evolving landscape of short-video platforms like TikTok and Douyin, user experience is defined by the seamless transition between videos. To achieve this, applications must "preload" content before a user even swipes. However, this creates a significant technical challenge: the preloading dilemma. If a system preloads too much and the user swipes away quickly, bandwidth is wasted—a massive cost at scale. If it preloads too little, the user encounters a rebuffering spinner, breaking the immersive experience.

DeLoad is a novel preloading framework designed to navigate this trade-off by moving away from rigid, fixed-size preloading and toward a dynamic, demand-driven approach. By modeling user behavior as a probability distribution and utilizing Deep Reinforcement Learning (DRL), DeLoad optimizes both bandwidth efficiency and Quality of Experience (QoE).

The Limitations of Traditional Preloading

Existing preloading strategies often suffer from three primary flaws:

  1. Bandwidth Inefficiency: Fixed-size data chunks often lead to "over-loading" for users who swipe rapidly, resulting in significant data waste.
  2. Rebuffering Risks: Static systems cannot adapt quickly enough to fluctuating network conditions, leading to playback interruptions.
  3. The Cold-Start Problem: Systems often struggle to predict the watch-time for new or "long-tail" videos that lack historical engagement data, leading to suboptimal preloading decisions.

The DeLoad Architecture

DeLoad introduces a three-pillar framework to address these challenges through sophisticated statistical modeling and machine learning.

1. Scalable Watch-Time Estimation (WTE)

Rather than attempting to predict a single, exact watch-time value—which is inherently uncertain—DeLoad models watch-time using the Weibull distribution. This statistical approach treats a user's "swipe-away" action as a system failure event, allowing the model to capture the probability of continued viewing at any given second. To ensure the system remains robust for new content, DeLoad employs a Multi-Dimensional Hybrid Framework that fuses specific video data, user history, and a "duration ladder" (a fallback mechanism based on video length) to estimate engagement for videos with no prior history.

2. Demand-Driven Selection

DeLoad introduces a new metric called Demand, defined as the probability that a specific, currently uncached portion of a video will be the next segment the user watches. The system maintains a global view of the current playlist and ranks all potential download tasks by their Demand score. This ensures that the most "urgent" data is always prioritized for the download queue.

3. DRL-Enhanced Dynamic Range

Once a video segment is selected for preloading, the system must decide exactly how much data to request. DeLoad utilizes a Deep Reinforcement Learning (DRL) agent, specifically the Proximal Policy Optimization (PPO) algorithm, to determine the optimal "range" or size of the download. This agent is deployed on the client-side (the user's device), allowing it to make millisecond-level decisions based on local network stability, device performance, and current buffer levels.

Experimental Results and Real-World Impact

The researchers evaluated DeLoad using a combination of massive offline datasets (3,000 network traces and 150,000 hours of video) and a large-scale online A/B test on Douyin, serving millions of active users.

Key findings from the production deployment include:

  • Bandwidth Savings: DeLoad reduced overall bandwidth consumption by 3.76%, representing a significant cost reduction for the platform.
  • Increased Engagement: Despite downloading less data on average, the system increased total user watch-time by 0.9‰, suggesting a smoother, more satisfying user experience.
  • QoE Improvements: In offline simulations, DeLoad outperformed baseline algorithms by 34.4% to 87.4% in composite QoE scores.
  • Device Optimization: The framework proved particularly effective for users on low-end devices, where intelligent buffer management is critical for maintaining playback fluidity.

Conclusion

DeLoad demonstrates that the key to efficient video delivery lies in embracing the uncertainty of human behavior. By treating watch-time as a probability distribution and delegating download sizing to a client-side DRL agent, platforms can simultaneously reduce infrastructure costs and improve user satisfaction. Future work will focus on further personalizing the models and optimizing the on-device machine learning SDK to ensure minimal impact on battery life and device performance.

Summary updated

BibTeX
@inproceedings{liu2025deload,
  author = {Tong Liu and Zhiwei Fan and Guanyan Peng and Haodan Zhang and Yucheng Zhang and Zhen Wang and Pengjin Xie and Liang Liu},
  title = {{DeLoad: Demand-Driven Short-Video Preloading with Scalable Watch-Time Estimation}},
  booktitle = {ACM Multimedia},
  year = {2025},
  doi = {10.1145/3746027.3754849},
}
Figure