Video QoE Inference with Machine Learning

HTTP adaptive streaming (HAS) has become the de-facto standard for delivering video over the Internet. More content providers like YouTube and Twitch have started generating and delivering high quality streams (usually 4k resolution) with advanced end-to-end encryption mechanisms. This huge increase in HAS encrypted traffic, creates a significant challenge for network providers in understanding what is happening on their infrastructures which limits their ability to manage network infrastructures properly. Due to such invisibility, the network providers could not take appropriate decisions for better optimizations, resulting in significant revenue lost. Inferring the quality of experience (QoE) of HAS-based streaming video services is important, but recent studies highlight that most of existing solutions that rely on packet inspections, showing low performance in inference accuracy. To address this issue, we develop a machine learning powered system that infers QoE factors such as startup delay, rebuffering and selected quality, for encrypted on-demand HAS streaming video services. Our solution uses two data-driven techniques: Deep Self Organizing Map (DSOM) and Multi Layer Perceptron Backpropagation (MLPB), allowing efficient accuracy with low error in inferring QoE factors over several public video datasets, compared to some state-of-the-art approaches.