Building outlier detection ensembles by selective parameterization of heterogeneous methods

摘要

We address the problem of selecting members of ensembles for unsupervised outlier detection. The challenge here is to identify individually accurate but diverse members due to unsupervised nature of the problem. For this, we herein propose AnD-SELECT: Accurate-and-Diverse Selector, which considers a set of heterogeneous outlier detection methods at input and systematically selects accurate parameter variants i.e. parameterization of each type. Outlier detection methods in this input set are chosen such that they usually exhibit the characteristics of either progressive or regressive performance behavior with increasing parameter values. We then consider a wide range of parameter variants of each of these methods. From such homogeneous set of a method type, the objective is to select the more accurate parameterization-end, while avoiding selection of both the ends together due to above mentioned characteristics. Therefore, either a single accurate variant or a set of two variants showing explicit trade-off between accuracy and diversity, get selected. Evaluation on benchmark datasets shows notable performance improvement over existing selectors.(c) 2021 Elsevier B.V. All rights reserved.