Investigating the Quality of Different Self-Organizing Map Topologies for Complex Data

Abstract

Self-Organizing Maps (SOM) are useful tools for visualizing high dimensional data. However, conventional SOM suffer from the border effect. Therefore, Spherical Self-Organizing Maps (SSOM) have been developed to remove such negative effects. In this paper, we extend the topology of SSOM by reconstructing the neighbors to propose the concept of Concentric Spherical Self-Organizing Maps (CSSOM). The major improvement of CSSOM is that it allows using an arbitrary number of spheres and such a topology could be applied in analyzing sequential and time series data. We conducted experiments using these SOM topologies on several datasets. The display schemas and several measures for the quality of SOMs are discussed with the experimental results. The comparison of the results indicates that the quality of SOM is improved through using specified CSSOM depending on the characteristics of the dataset. Show more