Fast Spherical Self Organizing Map - Use of Indexed Geodesic Data Structure -

In order to remove the “border effect”, several spherical Self-Organizing Maps (SOM) based on the geodesic dome have been proposed. However, existing neighborhood searching methods on the geodesic dome are much more time-consuming than searching on the normal rectangular/hexagonal grid. In this paper, we present detailed descriptions of the algorithms used in training the Geodesic SOM (GeoSOM), which we previously proposed. Experimental results show that the GeoSOM runs almost at the same speed of the conventional 2D SOM and it also represents the data more accurately. In visualizing the GeoSOM, a 2D data map can be created by projecting the spherical SOM onto 2D the plane. The user is able to select any point of interest to be the center of the 2D map. No retraining is required when changing the center point.