Design of a machine learning application for anomaly detection in real-time series

Abstract

In nearly all enterprises, time series-connected problems are a day-to-day issue which we should know how to face. The question to solve, what will happen to our metrics in the future, can only be answered using sophisticated prediction methods, such as the ones that will be studied in this project. There are numerous algorithms that can be applied to nearly every possible situation but, depending on some factors such as required quality of the prediction, length of the forecasted period, time and associated cost; we should be capable of selecting the appropriate method to successfully fulfil our needs. Time series data is a series of data points indexed in time order. Therefore, data is organized around relatively deterministic timestamps which, compared to random samples, may contain additional information that will be tried to be extracted. The objective of this project is, firstly, to study and understand the properties of time series data, with the intention of being capable of generate accurate predictions using different machine learning techniques. Once this is achieved, these methods will be combined with more sophisticated predictive algorithms with the intention of detecting anomalies in real time series datasets. In order to make this research, the IoT Department of Telefonica I+D has facilitated anonymised data of real clients, which will be used for testing all the methods and techniques presented along this project.