Comparison of Different Calculations of the Density-Based Local Outlier Factor

摘要

In the paper, we propose several new density-based algorithms for outlier detection. We present the detailed synoptic theoretical analysis of the algorithms that compute the local outlier factor as a function of the densities of the neighborhood of the objects in a set of objects. Based on this analysis we propose a new calculation of the radius of the neighborhood and we create 66 algorithms to compute the outlier factor. All the algorithms are tested in the complex experiments to describe their basic and also specific characteristics. The results are presented and discussed. Intuitively it seems that the way how the radius of the neighborhood is calculated is important. This idea led to numerous modifications of this part of the algorithms, but on the basis of the experiments we demonstrate that these modifications have only little influence, and we describe which part of the algorithms influence the outlier score the most and we recommend three generally applicable algorithms with specific characteristics.