Uplink vs. Downlink: Machine Learning-Based Quality Prediction for HTTP Adaptive Video Streaming.

Area of Science:

• Computer Science
• Machine Learning
• Network Engineering

Background:

• Streaming video constitutes the majority of internet traffic, necessitating accurate quality assessments by providers.
• Current machine learning solutions for monitoring streaming quality often demand extensive datasets, posing a challenge for practical implementation.

Purpose of the Study:

• To develop and compare three machine learning models for estimating video playback behavior.
• To focus on creating a lightweight approach that minimizes data and feature requirements while maintaining high performance.

Main Methods:

• Analysis of over 13,000 YouTube video streaming sessions using data from the native YouTube mobile app.
• Development and comparison of three machine learning models to predict quality of experience metrics.
• Focus on uplink request information for model input.

Main Results:

• The study examined key quality of experience metrics including initial playback delay, video quality, quality changes, and rebuffering events.
• Machine learning models were trained and evaluated to estimate playback behavior.
• A lightweight approach was prioritized, aiming for state-of-the-art performance with reduced data and features.

Conclusions:

• Lightweight machine learning models can effectively estimate video streaming playback behavior.
• Minimizing data and feature requirements is feasible without sacrificing performance in quality of experience prediction.
• This approach offers a practical solution for internet providers and network operators to assess streaming quality efficiently.