About Problems with Finding Outliers in a Single-Class Problem

Abstract

Outlier detection is considered in a one-class classification problem. The challenges of detection are related to the choice of loss functions, optimization criteria, and transformations of various types of features. Examples are given from subject areas where such problems exist. The absence of clustering methods for outlier detection and the possibility of their unambiguous interpretation are justified. The diversity of analysis options is related to the presence of uncertainties in selecting clustering parameters. As additional domain knowledge, an analysis of the environment’s nature for outlier objects is provided. For this purpose, kernel density estimates of outliers are calculated over local areas of fixed size. A method is presented for determining the location of objects closest to the class center. The task of analyzing meteorological data for the city of Tashkent has been solved. To enable prompt response to environmental violations, a transition to solving a two-class classification problem is proposed. It is assumed that one class describes an acceptable state, while the other represents a deviation from environmental standards.