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

2025-10-21

AI Summary

To address two key bottlenecks in short-video preloading—poor adaptability to dynamic task scale and scalability challenges in viewing duration prediction—this paper proposes DeLoad, a novel preloading framework. Methodologically, DeLoad introduces (1) a dynamic task chunking mechanism that adjusts download granularity in real time based on network conditions and user behavior; (2) a lightweight, multi-dimensional viewing duration prediction model balancing accuracy and deployability; and (3) a deep reinforcement learning–based adaptive preloading policy that jointly optimizes Quality of Experience (QoE) and bandwidth efficiency. Evaluated on a large-scale commercial platform, DeLoad achieves a 0.09% increase in total user viewing time, a significant reduction in stalling rate, a 3.76% decrease in bandwidth consumption, and improves the QoE metric from 34.4% to 87.4%.

Technology Category

• Machine Learning: Online Learning & Bandits
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

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.

Problem

• Optimizing preloading strategy for short-video streaming QoE
• Addressing limitations in dynamic download task sizing
• Developing scalable watch-time estimation for commercial deployment

Innovation

• Dynamic task sizing adapts to varying network conditions
• Multi-dimensional watch time estimation enhances prediction accuracy
• Deep reinforcement learning optimizes adaptive download decisions