Improvement of visual stability by adjusting initial map of SOM and using hierarchical clustering

MIYOSHI Tsutomu, MOMOI Shinji
Department of Media Informatics, Ryukoku University, Shiga, JAPAN

Abstract
SOM learning is influenced by the sequence of learning data and the initial feature map. In conventional method, initial value of feature map has set at random, so a different mapping appears even by same input data, so different impressions could be increased to the same data in different diagnosis. In this paper, we focused on visual stability of SOM feature map, and we proposed new initialization methods of SOM feature map and new map clustering method using hierarchical clustering. By experiments, proposed methods are visually stable than conventional method in the view of node mapping.

Keywords: self-organizing map, feature maps, visual stability, improvement method, clustering.