Anomalous assemblage detection using nearest neighbor distance

Abstract

The outlierness of an instance in this thesis is defined based on the distance between two instances. For some datasets, outliers may not be isolated and formed small clusters. C-anomalous assemblage is a group of associated outliers having the number of instances less than or equal to C percent of the total instances. This thesis presents the anomalous assemblage detection algorithm called CND using a nearest neighbor distance for an anomalous score. The algorithm computes the index k equal to floor function of C percent times the total number of instances and uses the k-nearest neighbor distance for representing an anomalous score. Then, the adjusted boxplot based on medcouple for skew distribution is used to generate the threshold for detecting outliers. The performance of CND is tested on two types of datasets which are synthetic and real-world datasets from UCI website comparing with WOF and LOF. The experimental results show that CND is better than WOF and LOF on datasets based on precision, recall, and F1-measure.