A Pattern Classifying System Based on the Coverage Regions of Objects

Abstract

A new statistical pattern classifying system is proposed to solve the problem of the "peaking phenomenon". In this phenomenon, the accuracy of a pattern classifier peaks as the features increase under a fixed size of training samples. Instead of estimating the distribution of class objects, the system generates a region on the feature space, in which a certain rate of class objects is included. The pattern classifier identifies the class if the object belongs to only one class of the coverage region, but answers "unable to detect" if the object belongs to the coverage region of more than one class or belongs to none. Here, the coverage region is simply produced from the coverage regions of each feature and then extended if necessary. Unlike the Naive-Bayes classifier, the independence of each feature is not assumed. In tests of the system on the classification of characters, the performance does not significantly decrease as the features increase unless apparently useless features are added.