A Graph Partitioning Approach to SOM Clustering

Abstract

Determining the number of clusters has been one of the most difficult problems in data clustering. The Self-Organizing Map (SOM) has been widely used for data visualization and clustering. The SOM can reduce the complexity in terms of computation and noise of input patterns. However, post processing steps are needed to extract the real data structure learnt by the map. One approach is to use other algorithm, such as K-means, to cluster neurons. Finding the best value of K can be aided by using an cluster validity index. On the other hand, graph-based clustering has been used for cluster analysis. This paper addresses an alternative methodology using graph theory for SOM clustering. The Davies-Bouldin index is used as a cluster validity to analyze inconsistent neighboring relations between neurons. The result is a segmented map, which indicates the number of clusters as well as the labeled neurons. This approach is compared with the traditional approach using K-means. The experimental results using the approach addressed here with three different databases presented consistent results of the expected number of clusters.