End to end encryption of internet traffic is getting more and more popular and common nowadays. Additionally, a large portion of this traffic consists of high resolution videos, by OTT video service providers, such as Youtube. This poses a problem to the network operators, especially mobile network operators, who want to maintain a high service quality, because their knowledge of the quality of videos delivered through their network is limited. The service providers also aim for a high quality service, and good Quality of Experience for the users, but there are matters that they can not control, such as network dimensioning and configurations of the radio access networks, that are the responsibilities of the (mobile) network operators.
There are multiple approaches of solving this problem, analytical, session modeling and machine learning based methods, that mostly try to calculate the Quality of Experience class or Mean Opinion Score associated with the video sessions. The thesis describes a machine learning based approach, to estimate the QoE, by estimating one of the most important influencing factors of the QoE, stalling - whether playing the video was continuous, or did it freeze unexpectedly. The solution provides this information on a fine granularity, calculating the probability of stalling for every second of a video session, based on the network traffic, without involvement of the end users. This provides the opportunity for root cause analysis, which is an important part of mitigating the problems.
The solution is based on a Long-Short Term Memory (LSTM) neural network, and achieves promising results on a small dataset. The exact configuration parameters will have to be verified on a larger dataset, but it nevertheless proves, that the approach is working, and solves a problem, that did not have such a detailed answer before.