Anomaly detection on time series from furthest neighbor window subseries

Abstract

Anomaly detection in time series is classified into three types which are point anomaly, contextual anomaly, and collective anomaly. This work proposes a novel method called the Furthest Neighbor Window Subseries (FNWS) for detecting contextual anomalies which normally appear in a time series dataset. Three quartiles representing a local distribution are computed and relocated by subtracting the first data point in the window subseries. A vector of three quartiles —the lower quartile, the median and the upper quartile —is used to compute the distances among all window subseries and the furthest k-nearest neighbor distance is picked as the score. The collection of the one-dimensional score is sorted and the score quartiles are computed. The interquartile range rule from the adjusted boxplot for skew distributions is applied to identify anomalies. The empirical experiments on the benchmark time series datasets from Yahoo with a list of labeled outliers are performed and evaluated using precision, recall, and F-measure. The results show that FNWS works effectively and accurately having the average scores more than 80% on all metrics.