Model Combination Methods for Outlier Ensembles

Abstract

An important part of the process of creating outlier ensembles is to combine the outputs of different detectors. The precise method for model combination has a significant impact on the effectiveness of a particular outlier detection method because of the varying theoretical effects of different combination methods. For example, the impact of the scheme of averaging is quite different from that of maximization in terms of the bias and variance of the result. Therefore, the choice of model combination has a crucial effect on the results of the ensemble.