A direct method for cluster analysis

摘要

An algorithm based on a least-mean-square (LMS) criterion is presented. This algorithm partitions a multi-dimensional data set directly into a desired number of clusters. The result is compared favorably to existing methods in both performance and computational efficiency. An efficient method for determining a reasonable set of distributed initial cluster centers based on principal component analysis is also presented. This clustering algorithm is shown to converge to a unique minimum based on the LMS criterion and is demonstrated by digital computer simulation and applied to the analysis of vectorcardiograms.