Anomaly prediction in mobile networks : A data driven approach for machine learning algorithm selection

In this paper, we propose a model for proactive anomaly detection in mobile networks. We show that when Key Performance Indicators (KPIs) are highly correlated, the linear regression gives a good accuracy in anomaly detection for a short prediction horizon. When the prediction horizon is far, KPIs become weakly correlated. We propose to transform discrete measurements into functional data and apply a functional data regression method to perform the prediction. We compare pro-posed models using KPI measurements obtained from a real Long Term Evolution (LTE) network. We show that an improvement in prediction performance can be obtained by using functional data analysis.