Two-phase outlier detection in multivariate time series

In this paper, an efficient two-phase algorithm for detecting outlying samples in multivariate time series (MTS) datasets is proposed. The Bounded Coordinate System (BCS) metric is used to measure the similarity between two MTS samples, and the outlierness of a sample is measured by average distance to its k nearest neighbors. We partition the data into clusters, and then use nested loop algorithm to find top-n outliers. A heuristic and two pruning rules are utilized to quickly remove MTS samples that are not possible outlier candidates, reducing significantly the distance computation among objects. Experiments on two real-world datasets show the efficiency of the proposed algorithm.