Applying Clustering to the Classification Problem

Abstract

The minimum-distance classifier (Duda & Hart, 1973) learns a single mean prototype for each class and uses a nearest neighbor approach for classification. A problem arises when classes cannot be accurately represented using a single prototype; multiple prototypes may be necessary. Our approach is to find groups of examples for each of the classes, generalize these groups into prototypes using a mean representation, and then classify using a nearest neighbor approach. K-means clustering is applied in unsupervised environments for finding groupings of examples. The problem with k-means clustering is finding the correct num