A Method for Comparing Self-organizing Maps: Case Studies of Banking and Linguistic Data

The method of self-organizing maps (SOM) is a method of exploratory data analysis used for clustering and projecting multi-dimensional data into a lower-dimensional space to reveal hidden structure of the data. The algorithm used retains local similarity and neighborhood relations between the data items. In some cases we have to compare the structure of data items visualized on two or more self-organizing maps (i.e. the information about the same set of data items is gathered in different tasks, from different respondents or using time intervals (lags)). In this paper we introduce a method for systematic comparison of SOM maps in the form of similarity measurement. Based on the idea that the SOM retains local similarity relations of data items those maps can be compared in terms of corresponding neighborhood relations. We give two examples of case studies and discuss the method and its applicability as an additional and more precise measure of similarity of SOM maps.