engleski

A Machine Learning Approach to Classifying YouTube QoE Based on Encrypted Network Traffic

Due to the widespread use of encryption in Over-The-Top video streaming traffic, network operators generally lack insight into application-level quality indicators (e.g., video quality levels, buffer underruns, stalling duration). They are thus faced with the challenge of finding solutions for monitoring service performance and estimating customer Quality of Experience (QoE) degradations based solely on passive monitoring solutions deployed within their network. We address this challenge by considering the concrete case of YouTube, whereby we present a methodology for the classification of end users' QoE when watching YouTube videos, based only on statistical properties of encrypted network traffic. We have developed a system called YouQ which includes tools for monitoring and analysis of application-level quality indicators and corresponding traffic traces. Collected data is then used for the development of machine learning models for QoE classification based on computed traffic features per video session. To test the YouQ system and methodology, we collected a dataset corresponding to 1060 different YouTube videos streamed across 39 different bandwidth scenarios, and tested various classification models. Classification accuracy was found to be up to 84% when using three QoE classes ("low", "medium" or "high") and up to 91% when using binary classification (classes "low" and "high"). To improve the models in the future, we discuss why and when prediction errors occur. Moreover, we have analysed YouTube's adaptation algorithm, thus providing valuable insight into the logic behind the quality level selection strategy, which may also be of interest in improving future QoE estimation algorithms.

Quality of Experience ; Video Streaming ; HTTP adaptive streaming ; YouTube ; network measurements ; passive monitoring ; machine learning

nije evidentirano