A METHOD TO ESTIMATE THE STATISTICAL CONFIDENCE OF CLUSTER SEPARATION

Cluster analysis contains several multivariate methods for the separat ion of patterns (clusters). The definition of the optimum or universal ly best cluster analysis is an unresolved issue. Three methods are of special importance: 1. The statistical confidence of cluster separatio n. 2. The definition of the optimal number of clusters. 3. The descrip tion of the internal cluster structure. Two new methods addressing the se problems are presented. On the basis of non-hierarchical minimum-di stance cluster analysis a new method is described that allows a separa tion of clusters in a statistically well-founded way. This method solv es problems one and two. Using a newly developed special rank-sum anal ysis, a solution to the third problem is possible. An example shows th e practicability of the proposed procedures.