Quality of Outlier Detection

The detection of outliers is one of the most vital issues in modern-day data analysis and exploration. Contrary to classification and clustering, there is no natural quality (performance) index in outlier detection and the primary aim of the research described in this paper is to create such an index. This will not only allow for the evaluation of the results of outlier detection algorithms but also the optimization of the parameter values or other quantities present in such procedures. The quality index investigated, performs particularly well with respect to frequency (probabilistic) types of approaches, notably in cases with substantial noise. The application of a nonparametric concept when constructing this index makes the proposed method practically independent of the examined dataset distribution. It can also be effectively used for multidimensional and multimodal problems.