U*C: Self-organized Clustering with Emergent Feature Maps

Alfred Ultsch

Abstract

A new clustering algorithm based on grid projections is proposed. This algorithm, called U*C, uses distance information together with density structures. The number of clusters is determined automatically. The validity of the clusters found can be judged by the U*-Matrix visualization on top of the grid. A U*-Matrix gives a combined visualization of distance and density structures of a high dimensional data set. For a set of critical clustering problems it is demonstrated that U*C clustering is superior to standard clustering algorithms such as K-means and hierarchical clusterings.