An Approach for Anomaly Detection & Prediction in Time-series Telecommunication Data

Anomaly detection is necessary in almost every domain these days. Being able to predict the occurrence of an anomaly ahead of time would also prove to be a good asset for many fields. Especially in Telecommunication, it is like a boon, as using these approaches helps to narrow down the problems. Also, if they know that an anomaly might occur, it will help them be prepared to solve the situation. This paper proposes an approach for the detection and prediction of anomalies in time-series telecom data. In this approach, anomaly detection is done using STL Decomposition, and the residue or deviations that are maximum are classified as anomalies. Then, using the whole dataset, a few data points are forecasted. Next, anomalies are predicted in this newly generated data, using XGBoost on the earlier labelled data. This work also presents a method to identify and alert the average frequency observed in the dataset.