Clustering Large, Multi-level Data Sets: An Apporach Based on Kohonen Self Organizing Maps

Standard clustering methods do not handle truly large data sets and fail to take into account multi-level data structures. This work outlines an approach to clustering that integrates the Kohonen Self Organizing Map (SOM) with other clustering methods. Moreover, in order to take into account multi-level structures, a statistical model is proposed, in which a mixture of distributions may have mixing coefficients depending on higher-level variables. Thus, in a first step, the SOM provides a substantial data reduction, whereby a variety of ascending and divisive clustering algorithms become accessible. As a second step, statistical modelling provides both a direct means to treat multi-level structures and a framework for model-based clustering. The interplay of these two steps is illustrated on an example of nutritional data from a multicenter study on nutrition and cancer, known as EPIC