Outlier detection in multivariate time series: an application of hybrid DNN-DBSCAN technique

Abstract

In spite of on-going advances and utilization of Deep Neural Networks (DNN) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) techniques to enhance outlier detection in multivariate time series (MTS) data, research is yet to explore an approach that integrates the capabilities of the two techniques for complex data representation. The paper therefore combines the two techniques to obtain the hybrid DNN-DBSCAN technique. It is demonstrated in simulated data that the resulting technique achieves improved precision and recall of outlier detection based on a number of performance metrics. In particular, the new procedure adequately captures the complexity involved with the underlying high-dimensionality of MTS data which poses problems for outlier detection to traditional methods.