Distance Boxplot for Unsupervised Outlier Detection

Anomaly detection is one of the key step in data preprocessing. Moreover, it is also vital in decision processing with applications such as intrusion detection and spam filtering. In this paper, an unsupervised anomaly detection method based on boxplot approach is proposed. In the proposed approach, a distance metric is integrated with the classical boxplot which results in a powerful hybrid approach capable of detecting both, either a single lying outlier or a cluster of outliers. Integration of distance metric takes into account the neighborhood estimation along with the density estimation. The concept realizes the compactness of data along with the position of underlying data points. This makes it suitable even for detecting small clusters of outlying data points which can be ignored as inlier data points because of their existence in a close proximity and high density. The usefulness of the proposed approach is demonstrated by some complex synthetic datasets.