Outlier detection with explanation facility

Abstract

We propose a simple and efficient method to detect exceptional data, which includes a novel end user explanation facility. After various designs, the best was based on an unsupervised learning schema, which uses an adaptation of the artificial neural network paradigm ART for the cluster task. In our method, the cluster that contains the smaller number of instances is considered as outlier data. The method provides an explanation to the end user about why this cluster is exceptional with regard to the data universe. The proposed method has been tested and compared successfully not only with well-known academic data, but also with a real and very large financial database that contains attributes with numerical and categorical values. © 2009 Springer Berlin Heidelberg.