Visualization of decision borders on spherical SOM and its Application to Clustering

AbstractIn this paper, we propose two methods for determining class borders. One approximates the borders on a SOM from the nearest neighbor datasets of different classes. The other especially approximates the borders on a spherical SOM using the coordinate system of a polygon surface. Both methods decide the border on the SOM using the characteristics of the SOM which can map a high dimensional dataset onto a low dimensional map which is usually 2 or 3 dimensional. Using the iris dataset of Fisher and the wine dataset, it is shown that the proposed methods allow the class borders to be successfully visualized, and in a comprehensible manner, and that the methods can improve the accuracy of a cluster analysis. The verification of the decision border, computed with one of the proposed methods, was performed with the spherical SOM and the dendrogram. The advantages of visualization were successfully demonstrated using two benchmark databases.