Time series anomaly detection based on DBSCAN ensembles

The quality of anomaly detection algorithms highly depends on the input parameters and internal structure of dataset, in addition this problem usually occurred in unsupervised setting leading to the conceptual complexity of quality measurement. In practice there is a significant variance of results of anomaly detection due to the huge amount of datasets under consideration having diverse internal structure. Outlier ensemble is a kind of technique which can improve the variance of anomaly detection and increase the overall quality of identification. In this paper we investigate the problem of anomaly detection in time series in unsupervised setting, propose the method of outlier ensemble construction based on DBSCAN algorithm, which uses the time series internal structure for adaptive input parameters selection. Experiments on synthetic and real datasets show the decrease of variance and high quality of method compared to popular techniques such as Median Absolute Deviation, One Class SVM, Isolation Forest, Local Outlier Factor and simple DBSCAN.