A self-organizing map for analysis of high-dimensional feature spaces with clusters of highly differing feature density

摘要

Self-Organizing Feature Maps (SOFM) facilitate a reduction and cluster analysis of high-dimensional fearture spaces. One property of these artificial neural nets is a smoothing of the input vectors and thus a certain insensitivity to outliers and clusters of low feature density. While classifying clusters of highly different feature density this property is undesirable. This paper introduces an algorithm which makes a projection of low feature density clusters onto a SOFM possible, too. The resulting weight vectors represent reference vectors of all clusters.