Learning Outlier Ensembles: The Best of Both Worlds – Supervised and Unsupervised
Barbora Micenková, Brian McWilliams, Ira Assent · 2014
Years of research in unsupervised outlier detection have pro-duced numerous algorithms to score data according to their exceptionality. However, the nature of outliers heavily de-pends on the application context and different algorithms are sensitive to outliers of different nature. This makes it very difficult to assess suitability of a particular algorithm without a priori knowledge. On the other hand, in many ap-plications, some examples of outliers exist or can be obtained in addition to the vast amount of unlabeled data. Unfortu-nately, this extra knowledge cannot be simply incorporated into the existing unsupervised algorithms. In this paper, we show how to use powerful machine learn-ing approaches to combine labeled examples together with arbitrary unsupervised outlier scoring algorithms. We aim to get the best out of the two worlds—supervised and un-supervised. Our approach is also a viable solution to the recent problem of outlier ensemble selection.