Cluster validity methods

Abstract

Clustering is an unsupervised process since there are no predefined classes and no examples that would indicate grouping properties in the data set. The majority of the clustering algorithms behave differently depending on the features of the data set and the initial assumptions for defining groups. Therefore, in most applications the resulting clustering scheme requires some sort of evaluation as regards its validity. Evaluating and assessing the results of a clustering algorithm is the main subject of cluster validity. In this paper we present a review of the clustering validity and methods. More specifically, Part I of the paper discusses the cluster validity approaches based on external and internal criteria.

Estimation of the Number of Clusters Using Multiple Clustering Validity Indices

“…Given a set of alternative clustering solutions, the preferred one is found by comparing the CVIs calculated for several candidate clustering solutions [8,16,17]. In order to find the number of clusters, a clustering algorithm is used with the number of clusters equal to k ∈ Ω, where Ω is the ordered set of candidate cluster numbers.…”

Cluster-wise assessment of cluster stability

“…Some alternative methods of cluster validation are homogeneity and/or separationbased validation indexes, comparison of different clustering methods on the same data, visual cluster validation, tests of homogeneity of the data set against a clustering alternative and use of external information, see Gordon (1999);Haldiki et al (2002);Hennig (2005);Milligan and Cooper (1985) and the references given therein.…”

Learning the Number of Clusters in Self Organizing Map

“…The single link technique is good at handling non-elliptical shapes, but is sensitive to noise and outliers. In this paper the quality of the clustering has been evaluated using external criteria (Overlap Indices Rand and Jaccard) frequently used (Halkidi et al, 2001(Halkidi et al, , 2002 :…”