A technique for unsupervised cluster recognition
William A. Porter · 1990
A methodology is presented for solving a basic problem in pattern recognition, that of unsupervised data clustering. Let X=(x/sub i/) denote a finite subset of the vector space S. Using a metric criterion, and with no prior information, determine the existence of clusters in the set X. Further, determine the number of clusters, the cluster membership of each point of X, and consequently a disjoint decomposition of X. Finally, if X is a moving window on a possibly infinite data stream, can these matters be resolved online? The design of the methodology given has features which are compatible with systolic array implementation and, in particular, VLSI technology. The performance of the design is evaluated using a simulation testbed. Some alternative techniques, based on SAS statistical software, are also explored and comparison with the resultant designs is given.>