Estimating QoE from QoS in real-time traffic: a Machine learning approach

As the title of thesis suggests, this research activity consists in analyzing data traces about video-conferencing tools and developing Machine-Learning approaches for estimating the Quality of Experience of a meeting through measurements on the data traffic. The knowledge of QoE is very valuable in the network providers’ point of view, since by knowing what is the quality experienced at the end user, they can improve the network performance accordingly. The main idea is to use a considerable number of data pcaps, collected from various types of meetings, in order to predict some target QoE metrics aiming to improve the meeting quality. So, basically by using measurements collected via Wireshark, a famous packet sniffer tool, at the vantage points of the network, we work on developing various ML approaches to map them into 3 main QoE targets, collected from application logs at the user side, which are: resolution, smoothness and concealment. In order to put the QoS measurements alongside with corresponding QoE values, a tool called Retina is used. Retina is a tool developed by SmartData research group of Politecnico Di Torino which produces a rich log of statistics on observed streams. Later on, once the file containing the dataset has been prepared by Retina, we use Jupyter Notebook web application tool to perform our ML approaches. The programming language being used in this research activity is Python and thanks to the built-in ML libraries in Jupyter Notebook we were able to easily perform various algorithms and techniques through some few lines of code. In this experimentation, there have been used various ML algorithms such as Decision Tree, Random Forest, XG Boost as well as different performance improvement techniques such as feature selection, outlier detection and treatment etc.