Unveiling YouTube QoE Over SATCOM Using Deep-Learning

Abstract

The importance of stored streaming video for current Internet traffic is undeniable, even in the context of satellite communications (SATCOM). Therefore, Internet service providers aim to deliver the highest quality of experience to their end users, although they are not able to assess it directly. Some machine learning techniques proposed in the literature have demonstrated their ability to predict the quality of experience based on traffic data analysis. However, these models cannot be directly applied in a SATCOM context without considering the specific characteristics of satellite links. Furthermore, some of them may not be suitable for real-time use. In this study, we monitored over 2,400 YouTube video sessions over an emulated satellite network to develop models capable of predicting the initial delay, played resolution, and stalling events. The collected dataset is available as an open source to the research community. The primary objective of this research is to provide a functional model for real-time applications. To achieve this, we reduced the required feature set to minimize computation time and resources, enabling a practical real-time implementation of the model while assessing its feasibility. We show that we successfully achieved a substantial reduction in the number of features while also observing a relative improvement in prediction. Our results yield prediction performance comparable to that of other studies on terrestrial networks. Using the reduced feature set, we present a real-time implementation and confirm the real-time viability of our work.

Keywords

• Deep learning
• HTTP adaptive video streaming
• machine learning
• network monitoring
• QoE
• SATCOM