A new pattern clustering method based on the Karhunen-Loeve expansion and moment-preserving thresholding is proposed. For a given set of unlabeled, nonparametric patterns with the number of clusters unknown a priori, if the patterns in each cluster are densely populated and the region between any two clusters is sparsely populated, then the proposed method can be employed to classify the patterns into clusters. The method consists basically of the use of the Karhunen-Loeve expansion to determine the directions for successive cluster analysis steps, the application of moment-preserving thresholding to find the major portion of a given pattern set, and a split-and-merge process for cluster construction. The phenomenon of cluster overlapping in any direction is overcome by the idea of successively shrinking an approximation of the major portion of the pattern set, called central region, within which splitting is performed. Examples are included to show the effectiveness of the method.
Relation:
Proceedings of 1987 National Computer Symposium, Taipei, Taiwan, Republic of China
