A Modified Clustering Method Based on Self-Organizing Maps and Its Applications
Le Yang, Zhongbin Ouyang, Yong Shi · Procedia Computer Science · 2012
Self-organizing map (SOM) is one of the most popular neural network methods for cluster analysis. Clustering methods using SOM usually are two-stage procedures: first original data are projected onto a set of prototypes on an ordered grid by SOM, and these prototypes can be seen as proto-clusters which will be grouped in the second stage to obtain finally clustering results. Many methods have been proposed to cluster the proto-clusters, among which the prototypes are considered as isolated vectors, without relationship of SOM; others are based on the U-matrix, which represents the local distance structure coming from the topology of SOM. In this paper, we propose a novel method more related to SOM for the purpose to cluster the proto-clusters. In the second stage we use the grid information alternatively, regarding it as a graph partitioned by graph cut algorithm well-known as Normalized cut. We apply this method on image processing and seismic data analysis and obtain reasonable results.