Growing Hierarchical Self-Organizing Map Using Category Utility

Abstract

In order to automatically obtain hierarchical knowledge representation from a certain data, an unsupervised learning method has been developed that overcomes two problems of the growing hierarchical self-organizing map (GHSOM) method, which uses the quantization error, the deviation of the input data, as evaluation measure of the growing maps: proper control of the growth process of each map is difficult due to the use of the quantization error and the clusters in the hierarchical structure may be excessively subdivided. This improved GHSOM method uses the category utility (CU), a measure used in conceptual clustering for predicting the preferred level of categorization, instead of the quantization error. The CU is useful for organizing the clustering so that people can effortlessly understand it. The basic principle of this method is that the growth and unification processes are appropriately and autonomously controlled by the CU. Evaluation using computer experiments showed that the proposed method can automatically construct an appropriate hierarchical and topological knowledge representation for high-dimensional input data through unsupervised learning. It also showed that it is easier to use and more effective than the original conventional GHSOM method using the quantization error as an evaluation measure.