Data analysis of not well separable clusters of different feature density with a two-layer classification system comprised of a SOM and an ART 2-A network

This paper introduces a two-layer classification system. The system is suitable for classification tasks of high-dimensional feature spaces which contain not well separable clusters of different feature density. The first layer, a modified SOM, calculates a set of reference vectors of the feature distribution under preservation of neighborhood relations. The modification supports the learning of definite neurons into the direction of clusters with low feature density better than the basic algorithm. In the second layer an ART 2-A network classifies similar and possibly scaled reference vectors into the same class. After classification, each class can contain several reference vectors, which characterize a distribution density function inside the determined classes. In addition an application of the two-layer classification system in the field of biomedical data analysis is described.