Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

TL;DR

This paper tackles the challenge of preloading for short-video platforms by introducing DeLoad, a demand-driven framework that combines dynamic task sizing, Weibull-based watch-time distribution modeling, and a DRL-enhanced range controller. It integrates server-side watch-time estimation with an on-device decision engine that uses a Demand metric to prioritize downloads and a continuous action space to determine range durations. Offline trace-based evaluations show substantial QoE improvements (34.4%–87.4%) and reduced rebuffering, while large-scale online A/B tests report a modest but meaningful increase in watch time and bandwidth savings, with deployment on Douyin validating practical impact. The work demonstrates how dynamic range granularity and robust watch-time prediction can deliver significant QoE gains and bandwidth efficiency in production short-video systems.

Abstract

Short video streaming has become a dominant paradigm in digital media, characterized by rapid swiping interactions and diverse media content. A key technical challenge is designing an effective preloading strategy that dynamically selects and prioritizes download tasks from an evolving playlist, balancing Quality of Experience (QoE) and bandwidth efficiency under practical commercial constraints. However, real world analysis reveals critical limitations of existing approaches: (1) insufficient adaptation of download task sizes to dynamic conditions, and (2) watch time prediction models that are difficult to deploy reliably at scale. In this paper, we propose DeLoad, a novel preloading framework that addresses these issues by introducing dynamic task sizing and a practical, multi dimensional watch time estimation method. Additionally, a Deep Reinforcement Learning (DRL) enhanced agent is trained to optimize the download range decisions adaptively. Extensive evaluations conducted on an offline testing platform, leveraging massive real world network data, demonstrate that DeLoad achieves significant improvements in QoE metrics (34.4% to 87.4% gain). Furthermore, after deployment on a large scale commercial short video platform, DeLoad has increased overall user watch time by 0.09% while simultaneously reducing rebuffering events and 3.76% bandwidth consumption.

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

TL;DR

This paper tackles the challenge of preloading for short-video platforms by introducing DeLoad, a demand-driven framework that combines dynamic task sizing, Weibull-based watch-time distribution modeling, and a DRL-enhanced range controller. It integrates server-side watch-time estimation with an on-device decision engine that uses a Demand metric to prioritize downloads and a continuous action space to determine range durations. Offline trace-based evaluations show substantial QoE improvements (34.4%–87.4%) and reduced rebuffering, while large-scale online A/B tests report a modest but meaningful increase in watch time and bandwidth savings, with deployment on Douyin validating practical impact. The work demonstrates how dynamic range granularity and robust watch-time prediction can deliver significant QoE gains and bandwidth efficiency in production short-video systems.

Abstract

Short video streaming has become a dominant paradigm in digital media, characterized by rapid swiping interactions and diverse media content. A key technical challenge is designing an effective preloading strategy that dynamically selects and prioritizes download tasks from an evolving playlist, balancing Quality of Experience (QoE) and bandwidth efficiency under practical commercial constraints. However, real world analysis reveals critical limitations of existing approaches: (1) insufficient adaptation of download task sizes to dynamic conditions, and (2) watch time prediction models that are difficult to deploy reliably at scale. In this paper, we propose DeLoad, a novel preloading framework that addresses these issues by introducing dynamic task sizing and a practical, multi dimensional watch time estimation method. Additionally, a Deep Reinforcement Learning (DRL) enhanced agent is trained to optimize the download range decisions adaptively. Extensive evaluations conducted on an offline testing platform, leveraging massive real world network data, demonstrate that DeLoad achieves significant improvements in QoE metrics (34.4% to 87.4% gain). Furthermore, after deployment on a large scale commercial short video platform, DeLoad has increased overall user watch time by 0.09% while simultaneously reducing rebuffering events and 3.76% bandwidth consumption.
Paper Structure (35 sections, 7 equations, 6 figures)

This paper contains 35 sections, 7 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Lack of Adaptation to Dynamic Range
  4. Range request granularity directly affects throughput, thus impacting playback smoothness
  5. Range request granularity directly affects scheduling frequency, thus impacting playback smoothness
  6. Range request granularity affects bandwidth control, thus impacting bandwidth cost
  7. Inaccurate and Non-Universal Watch-Time Estimation
  8. Prediction Accuracy.
  9. Universality for Large-Scale Platforms.
  10. Design
  11. System Overview
  12. Watch-time Distribution Prediction
  13. Weibull modeling.
  14. Joint Multi-Dimensional Estimation.
  15. Preloading Video Selection
  16. ...and 20 more sections

Figures (6)

  • Figure 1: Impact analysis under different fixed range request durations. Top: Overall impact. Bottom: Comparative analysis between casual and persistent viewers.
  • Figure 2: End-to-End System Overview of DeLoad Short-Video Preloading Framework
  • Figure 3: Weibull Watch-Time Distribution Overview
  • Figure 4: Architecture of Dynamic Video Range Model
  • Figure 5: Offline Evaluation Results
  • ...and 1 more figures