A study on evaluation measures for unsupervised outlier detection

Outlier detection is a data analysis method based on data mining techniques and is used to identify outlying observations which might have significance in a dataset. Research on outlier detection, however, has mainly focused on supervised approaches, which require labeled training and test datasets. Unsupervised approaches are more appropriate for many applications such as network intrusion detection and fraud detection, but the suitability of these methods to determine the degree of outlierness of a dataset has not been fully addressed because the ground truth is usually unavailable. In this paper, evaluation measures for unsupervised outlier detection, which can effectively measure the outlierness of a dataset, are proposed. To verify the effectiveness of the proposed methods, experiments were conducted with University of California Irvine machine learning datasets using a k-nearest neighbors (k-NN) algorithm.