Data visualization and analysis with self-organizing maps in learning metrics

Abstract

High-dimensional data can be visualized and analyzed with the Self-Organizing Map, a method for clustering data and visualizing it on a lower-dimensional display. Results depend on the (often Euclidean) distance measure of the data space. We introduce an improved metric that emphasizes important local directions by measuring changes in an auxiliary, interesting property of the data points, for example their class. A Self-Organizing Map is computed in the new metric and used for visualizing and clustering the data. The trained map represents directions of highest relevance for the property of interest. In data analysis it is especially beneficial that the importance of the original data variables throughout the data space can be assessed and visualized. We apply the method to analyze the bankruptcy risk of Finnish enterprises.