Determination of Class Borders on the spherical SOM with Non-equal Class Distributions and its Visualization

Abstract

We propose an approach to the determination of class borders on a SOM with non-equal class distributions. Our approach treats the class distribution as a variance-covariance matrix. The class distribution is expressed by a variance-covariance matrix and its decision border between the classes is determined from input data by using the eigenvalues and the corresponding eigenvectors of the matrices. Using the iris dataset of Fisher, it is shown that our approach allows the effect of non-equal class distributions on the decision borders to be successfully visualized in a qualitative and comprehensible manner.