Detecting Anomalies in a Time Series Database

Abstract

We present a comprehensive evaluation of a large number of semi-supervised anomaly detection techniques for time series data. Some of these are existing techniques and some are adaptations that have never been tried before. For example, we adapt the window based discord detection technique to solve this problem. We also investigate several techniques that detect anomalies in discrete sequences, by discretizing the time series data. We evaluate these techniques on a large variety of data sets obtained from a broad spectrum of application domains. The data sets have different characteristics in terms of the nature of normal time series and the nature of anomalous time series. We evaluate the techniques on different metrics, such as accuracy in detecting the anomalous time series, sensitivity to parameters, and computational complexity, and provide useful insights regarding the effectiveness of different techniques based on the experimental evaluation.