| Self-organizing maps (SOM) were developed by Kohonen in late 1980's in response to the growing amount, of large data sets that represented a problem for thorough analysis with standard tools and techniques. The SOM approach provides clustering capabilities for multi-dimensional data sets, and a variety of implementations provide for comparison of these maps to other visualizations.;The purpose of this work is to develop a constrained SOM-based unsupervised clustering algorithm (SOMn) for interactive analysis of large data sets. We believe that the power of SOM has not been fully exploited, and there is room for better interpretation as well as visualization of map outputs.;We first provide a link between the constrained unsupervised SOM and existing two- and three-dimensional visualization techniques, exhibiting scatter plot to SOM visualization interpolation utilizing variable neighborhood sizes. The SOMn frame-work for visualization display utilizes the SOM approach, binning, and coarse and fine mapping.;We extend the SOMn approach to address the occlusion problem, providing a novel jittering and focus and context technique. Exploration was continued by providing intrinsic cluster dimensionality metric and intrinsic SOMn dimensionality that assess the technique and analyzes the display of data. We develop lines of separation, a visual tool for identification of distances among neighboring output nodes, and directional vectors, both as data-driven tools that aid in the detection of clusters and outliers. |
