2D and 3D Neural-Network Based Visualization of High-Dimensional Biomedical Data

Abstract

In this paper we integrate self-organizing map algorithm (SOM) with scatter plot and Radviz, extending these visualizations into the third dimension and reducing overlap. Classic visualizations are used as the two- dimensional base, combined with a self-organizing map that extends them into the third dimension, with an adjusted neighborhood function. This approach solves the problem of overlap where more than one point plots to the same space and uncovers additional information about relationships inherent in high-dimensional data sets, including distribution of points, outliers and associations. Case studies are presented on a microarray and miRNA data sets.