Traffic Forecast in Mobile Networks: Classification System using Machine Learning

In this work, we propose a methodology to improve the precision of cell traffic forecasting with a machine learning approach. To develop this methodology, we first performed a systematic analysis in order to reduce bias by selecting the cells with less missing data occurrences. Then, we selected the features and trained a classifier to allocate the cells between predictable and non-predictable, taking into account previous traffic forecast error. The Naive Bayes classifier and Holt-Winters method was selected to perform the proposed methodology in real time. The system was applied to a set of 786 cells in a real network. The classifier presented a 91% accuracy, which leads the predictable cells, using Holt-Winters, to present an average RMSE of 2.74%. This means that it is now possible to implement optimisation algorithms that are highly sensitive to traffic prediction.