Class separation through variance: a new application of outlier detection

Abstract

This paper introduces a new outlier detection approach and discusses and extends a new concept, class separation through variance. We show that even for balanced and concentric classes differing only in variance, accumulating information about the outlierness of points in multiple subspaces leads to a ranking in which the classes naturally tend to separate. Exploiting this leads to a highly effective and efficient unsupervised class separation approach. Unlike typical outlier detection algorithms, this method can be applied beyond the ‘rare classes’ case with great success. The new algorithm FASTOUT introduces a number of novel features. It employs sampling of subspaces points and is highly efficient. It handles arbitrarily sized subspaces and converges to an optimal subspace size through the use of an objective function. In addition, two approaches are presented for automatically deriving the class of the data points from the ranking. Experiments show that FASTOUT typically outperforms other state-of-the-art outlier detection methods on high-dimensional data such as Feature Bagging, SOE1, LOF, ORCA and Robust Mahalanobis Distance, and competes even with the leading supervised classification methods for separating classes.