Learning from large-scale commercial networks: challenges and knowledge extraction towards machine learning use cases

Machine Learning (ML) algorithms are proposed to replace conventional algorithms in the area of wireless networking. Many of the suggested algorithms are often based on simulators or smallscale test-beds. We provide a study based on a dataset collected over a large commercial network, and highlight some of the real network dynamics that learning agents need to cope with. Our dataset includes not only measurements from the User Equipment (UE) but also integrates information from the network. Based on the collected data, we highlight some of the aspects that are important for the design of learning agents and discuss potential dataset characteristics that might hinder the learning process. Then we discuss what dataset characteristics can facilitate the deployment of ML algorithms in the real networks. Finally, we showcase how throughput prediction can be implemented by using ML techniques and provide some examples and insights on feature engineering and the training process.