Modeling Outlier Score Distributions

A common approach to outlier detection is to provide a ranked list of objects based on an estimated outlier score for each object. A major problem of such an approach is determining how many objects should be chosen as outlier from a ranked list. Other outlier detection methods, transform the outlier scores into probability values and then use a user-predefined threshold to identify outliers. Ad hoc threshold values, which are hard to justify, are often used. Outlier detection accuracy can be seriously reduced if an incorrect threshold value is used. To address these problems, we propose a formal approach to analyse the outlier scores in order to automatically discriminate between outliers and inliers. Specifically, we devise a probabilistic approach to model the score distributions of outlier scoring algorithms. The probability density function of the outlier scores is therefore estimated and the outlier objects are automatically identified.