TLDR: DeLoad is a new framework designed to improve short-video streaming by dynamically adjusting video preloading based on user demand and predicted watch-time. It addresses limitations of existing methods by introducing dynamic download task sizing and a scalable, multi-dimensional watch-time estimation model. A Deep Reinforcement Learning agent optimizes download decisions, leading to significant improvements in user experience (QoE) and a reduction in bandwidth consumption, as demonstrated by its deployment on a major commercial platform.
Short video platforms have become incredibly popular, dominating digital media with their rapid-fire content and quick user interactions. However, behind the seamless experience lies a significant technical challenge: how to effectively preload videos to ensure smooth playback while also being efficient with internet bandwidth. Researchers have been working on this problem, and a new framework called DeLoad offers a promising solution.
Existing methods for preloading short videos often fall short in two key areas. First, they struggle to adapt the size of download tasks to changing network conditions and user behavior. This means they might download too much or too little data, leading to either wasted bandwidth or frustrating interruptions. Second, their models for predicting how long a user will watch a video are often difficult to implement reliably at a large scale, especially for new or less popular content.
DeLoad addresses these issues by introducing a novel approach that combines dynamic task sizing with a practical, multi-dimensional method for estimating watch-time. The framework also incorporates a Deep Reinforcement Learning (DRL) agent that learns to optimize download decisions adaptively, making it smarter and more responsive to real-world conditions.
The DeLoad framework consists of three main components. The first is a robust watch-time modeling and prediction system. Instead of just guessing a single watch-time, DeLoad models the full probability distribution of how long a user might watch a video. It uses a statistical tool called the Weibull distribution, which is efficient and interpretable. To make these predictions scalable, especially for videos without much viewing history, DeLoad employs a multi-dimensional approach. It considers video-specific data, user-specific viewing habits, and a ‘duration ladder’ as a fallback, fusing these insights to create a comprehensive watch-time estimate.
The second component is demand-driven video selection. Based on the accurate watch-time predictions, DeLoad defines a metric called ‘Demand’ to prioritize which video to download next. Essentially, it calculates the probability that an unbuffered part of a video will be watched soonest among all videos in a user’s playlist. This intelligent prioritization ensures that the most relevant content is ready when the user wants it, balancing immediate playback needs with future viewing expectations.
Finally, DeLoad features a DRL-enhanced dynamic range selection model. This is where the framework truly shines in adapting download sizes. Unlike older systems that use fixed-size chunks, DeLoad’s DRL agent learns to determine the optimal download duration for each preloading task. This agent, deployed on the user’s device, continuously adapts to dynamic network conditions, ensuring that the system is always making the best decision for both playback quality and bandwidth efficiency. For more technical details, you can refer to the full research paper: DeLoad: Demand-Driven Short-Video Preloading with Scalable Watch-Time Estimation.
Extensive evaluations have shown DeLoad’s effectiveness. In offline tests using massive real-world network data, DeLoad achieved significant improvements in Quality of Experience (QoE) metrics, with gains ranging from 34.4% to 87.4%. More impressively, after being deployed on a large-scale commercial short-video platform (Douyin), DeLoad led to a 0.9‰ increase in overall user watch-time. At the same time, it successfully reduced rebuffering events and cut bandwidth consumption by 3.76%. These results highlight DeLoad’s ability to deliver a better user experience while also being economically beneficial for platforms.
Also Read:
- Justitia: Enhancing Performance and Equity in LLM Application Scheduling
- Unlocking Deeper Video Reasoning Through Entropy Control
The research demonstrates that dynamically adapting download task sizes and accurately predicting user watch-time are crucial for optimizing short-video streaming. DeLoad represents a significant step forward, offering a robust and practical framework that has already proven its value in a real-world commercial setting, serving millions of daily users.
