A machine learning application for latency prediction in operational 4G networks

Measuring performance on Internet is always challenging. When it comes to the mobile networks, the variety of technology characteristics coupled with the opaque network configuration make the performance evaluation even a more difficult task. Latency is one of the aspects having the largest impact on the performance and on the end users' Quality of Experience. In this paper, we present a machine learning approach that, exploiting real mobile network data on the end user, try to predict the latency in a real operational network. We consider a large-scale dataset with more than 238 million latency measurements coming from 3 different commercial mobile operators. The presented methodology flattens the RTT values into several bins, turning the latency prediction problem to a multi-label classification problem. Then, three well-known supervised algorithms are exploited to predict the latency. The obtained results highlight the importance of representative dataset from operational network. It calls for further improvements on the algorithm selection, tuning, and their predictive capabilities.