Calibration and Evaluation of Outlier Detection with Generated Data

Outlier detection is an essential part of data science --- an area with increasing relevance in a plethora of domains. While there already exist numerous approaches for the detection of outliers, some significant challenges remain relevant. Two prominent such challenges are that outliers are rare and not precisely defined. They both have serious consequences, especially on the calibration and evaluation of detection methods. This thesis is concerned with a possible way of dealing with these challenges: the generation of outliers. It discusses existing techniques for generating outliers but specifically also their use in tackling the mentioned challenges. In the literature, the topic of outlier generation seems to have only little general structure so far --- despite that many techniques were already proposed. Thus, the first contribution of this thesis is a unified and crisp description of the state-of-the-art in outlier generation and their usages. Given the variety of characteristics of the generated outliers and the variety of methods designed for the detection of real outliers, it becomes apparent that a comparison of detection performance should be more distinctive than state-of-the-art comparisons are. ... mehrSuch a distinctive comparison is tackled in the second central contribution of this thesis: a general process for the distinctive evaluation of outlier detection methods with generated data. The process developed in this thesis uses entirely artificial data in which the inliers are realistic representations of some real-world data and the outliers deviations from these inliers with specific characteristics. The realness of the inliers allows the generalization of performance evaluations to many other data domains. The carefully designed generation techniques for outliers allow insights on the effect of the characteristics of outliers. So-called hidden outliers represent a special type of outliers: they also depend on a set of selections of data attributes, i.e., a set of subspaces. Hidden outliers are only detectable in a particular set of subspaces. In the subspaces they are hidden from, they are not detectable. For outlier detection methods that make use of subspaces, hidden outliers are a blind-spot: if they hide from the subspaces, searched for outliers. Thus, hidden outliers are exciting to study, for the evaluation of detection methods that use subspaces in particular. The third central contribution of this thesis is a technique for the generation of hidden outliers. An analysis of the characteristics of such instances is featured as well. First, the concept of hidden outliers is broached theoretical for this analysis. Then the developed technique is also used to validate the theoretical findings in more realistic contexts. For example, to show that hidden outliers could appear in many real-world data sets. All in all, this dissertation gives the field of outlier generation needed structure and shows their usefulness in tackling prominent challenges of the outlier detection problem.