Identifying single good clusters in data sets

Abstract

Local patterns in the form of single clusters are of interest in various areas of data mining. However, since the intention of cluster analysis is a global partition of a data set into clusters, it is not suitable to identify single clusters in a large data set where the majority of the data can not be assigned to meaningful clusters. This paper presents a new objective function-based approach to identify a single good cluster in a data set making use of techniques known from prototype-based, noise and fuzzy clustering. The proposed method can either be applied in order to identify single clusters or to carry out a standard cluster analysis by finding clusters step by step and determining the number of clusters automatically in this way. © Springer-Verlag Berlin Heidelberg 2006.