Data driven evaluation of quality of service in mobile radio networks via machine learning

Design and implementation of AI algorithms for the evaluation of service quality in mobile radio networks. This thesis project aims to study the application of innovative paradigms such as artificial intelligence and machine learning within a highly dynamic and constantly evolving context, such as telecommunications. Mobile operators are constantly looking for innovative services to provide to their customers, while trying to optimize those already provided. Without doubt, the data service is central, as is shown by the numerous actions that aim to integrate it into numerous and varied fields (just think of the advent of the 5G and the IoT - Internet of Things). Among the possible areas of integration there are voice and messaging services. Currently, the technology foreseen for this objective is called VoLTE (Voice over LTE), which allows to establish voice calls on the LTE network based on a full IMS (IP Multimedia Subsystem) architectural model. However, nowadays this technology is not available everywhere, so the operators rely, at the moment, on an intermediate solution known as CSFB (Circuit Switched Fall-Back), which allows to switch (fall-back) from LTE network to legacy 3G / 2G networks to make a call or send a message, and then to return to the LTE network once the operation is completed. Mobile operators are interested in benchmark testing regarding the quality of the CSFB technology, in order to optimize their infrastructure where the need arises. This analysis operation is very laborious and takes a very long time. We tried to create artificial intelligence algorithms based on logics dictated by the experience of industry analysts, with the aim of automating these cataloguing operations. After obtaining a good accuracy of the results of the algorithms, compared to those of the analysts, we have passed to the application of the machine learning technique known as anomaly detection. This technique is generally used to identify objects, events or observations, which differ significantly from most of the data. Specifically, it was used to detect significant changes in failed mobile telephone calls in a specific geographical area, so that analysts, taking note of the variation, could carry out a more in-depth analysis. With this thesis elaboration we worked on two different aspects, on one hand the development of artificial intelligence algorithms for the automation of the analysis processes, while on the other the integration of specific machine learning techniques, such as the anomaly detection, on the data obtained from the algorithms. Therefore, the main objective is to provide an additional service to the analysts for the analysis of the benchmark logs. The tests carried out have demonstrated the potential of using these innovative techniques in this particular context, highlighting also some limitations that will be the object of possible future optimizations.