Abstract
Cluster extraction with Self-Organizing Map(SOM) can extract clusters of arbitrary distribution
shapes. However, like k-means method, the SOM-based clustering procedure may fail in cluster extraction when clusters have significantly different cluster-size or density each other. Recently, one technique called x-means which improve k-means method has been proposed. x-means method applies information criterion to k-means method and it can extract clusters based on a likelihood as a cluster. In this paper, we propose an application of information criterion to SOM-based clustering. This method is expected to have both advantages of SOM-based method and x-means method, that is, it supports clusters of arbitrary distribution shapes and likelihood-based cluster extraction. Through some experiments of clustering for several types of dataset, we confirmed that proposing method can extract clusters more correctly compared with k-means method.
