A memory adjustable software system for clustering and retrieval

This paper describes the structure and capabilities of a memory adjustable, menu driven software system, mostly written in C programming language for the classification and retrieval of multidimensional data. To group patterns, the system uses the single linkage clustering algorithm method, which requires the storage of a similarity matrix in the internal computer memory. Depending on the size of the data file, this requirement may limit the use of this method to computers with large memory. To overcome this problem, a new memory conservative single linkage clustering algorithm was developed and employed. In most existing systems, the user is required to declare the number of desired groupings. However, because the proper number of clusters significantly impacts the outcome of clusters, the system presented in this paper may determine the optimum number of clusters by calculating a measurement index. Finally, given a new pattern, the system determines its k most similar patterns in the data set by using an efficient branch and bound algorithm. The system is capable of graphing the results in any two dimensional plane.