An Automatic Cluster Analysis Algorithm.
Final rept. Jun 72-May 75,
AEROSPACE MEDICAL RESEARCH LAB WRIGHT-PATTERSON AFB OHIO
Pagination or Media Count:
This technical report presents an algorithm that finds clusters in a set of data. Examples of applications of this algorithm are the separation of targets from clutter in reconnaissance imagery and the determination of prototypes from any set of data represented in vector form, such as hand-printed letters, electrocardiographic data and electroencephalographic data. The approach used is to start with the closet vector to the data mean as the trial prototype for the first cluster. A hypersphere that is centered on the trial prototype is then defined. The radius of the hypersphere is incremented, and the prototype updated, until one of several termination criteria is met. The furthest vector from the resultant cluster is then used to start the search for the next cluster and the process is repeated until no new clusters can be located. The algorithm also merges clusters which are their own closest neighbors, are sufficiently close to each other, and result in a sufficiently small variance. These criteria were empirically derived and are related to data statistics. Author
- Statistics and Probability