Clustered Data Separation via Barycentric Radial Visualization

This paper addresses visualizing clusters of multi-dimensional data using barycenters as cluster representatives within the RadViz radial visualization technique. RadViz is a composition of two mappings. Where in this twostage mapping the cluster barycenters are formed is a key decision. Motivated by the nature of the second mapping, we form cluster barycenters at the end of the first stage, rather than at the start of the first stage. In the second stage we must select an appropriate configuration of dimensional representatives (dimensional anchors). Since this problem is intractable we present a heuristic to: 1) separate clusters and 2) move clusters away from the barycenter of the dimensional anchors. The heuristic uses our prior Voronoi quality assessment technique and our recent observation that circular motion of dimensional anchors confines each data image to an annulus. We demonstrate the benefit of our barycentric approach for a variety of clustered datasets.