TR-IIS-04-011    PDF format

Abstract

In this paper we present a series of data-reduction methods for classifying an unknown object as one member of a large set of possible patterns. The first reduction method is a learning algorithm that reduces a gigantic set of training samples into a condensed set of templates, each represented in vector form. When used in a testing process, these templates hold target patterns within the nearest K templates for almost all unknown objects, where K is a small number. The second reduction method exploits the nature of templates and classification trees to form a fast tree-retrieval mechanism. Experiment results show that this retrieval mechanism is a lot more effective than the K-means clustering method in meeting the same objective. The third method is a disambiguation method that supplies a condensed set of confusing pairs. This method exploits an effective binary classification technique to re-evaluate all the confusing pairs that appear in the nearest K templates for each unknown object, and thus improves the accuracy rate of the final classification decision.