Methodically Unified Procedures for Outlier Detection, Clustering and Classification

Abstract

In the practice of data analysis some problems for many-sided researches are caused by the methodological variety of specific algorithms, often leading to laborious interpretations and time-consuming studies. This paper presents the concept of methodically unified procedures, based on kernel estimators, for three fundamental tasks: outlier detection, clustering, and classification. Their clear interpretation facilitates the applications and potential individual modifications. The investigated procedures are distribution-free, enabling analysis and exploration of data with any distributions, also when elements are grouped in several separated parts. The results obtained depend not only on the values of particular attributes, but above all on the complex relationships between them.