Automatic optimization of outlier detection ensembles using a limited number of outlier examples
Niko Reunanen, Tomi Räty, Timo Lintonen
Abstract
Abstract In data analysis, outliers are deviating and unexpected observations. Outlier detection is important, because outliers can contain critical and interesting information. We propose an approach for optimizing outlier detection ensembles using a limited number of outlier examples. In our work, a limited number of outlier examples are defined as from 1 to 10% of the available outliers. The optimized outlier detection ensembles consist of outlier detection algorithms, which provide an outlier score and utilize adjustable parameters. The automatic optimization determines the parameter values, which enhance the discrimination of inliers and outliers. This increases the efficiency of the outlier detection. Outliers are rare by definition, which makes the optimization with a few examples beneficial. Obtaining examples of outliers can be prohibitively challenging, and the outlier examples should be used efficiently.