Subspace Shapes: Enhancing High-Dimensional Subspace Structures via Ambient Occlusion Shading

Abstract: We test the hypothesis whether transforming a data matrix into a 3D shaded surface or even a volumetric display can be more appealing to humans than a scatterplot since it makes direct use of the innate 3D scene understanding capabilities of the human visual system. We also test whether 3D shaded displays can add a significant amount of information to the visualization of high-dimensional data, especially when enhanced with proper tools to navigate the various 3D subspaces. Our experiments suggest that mainstream users prefer shaded displays over scatterplots for visual cluster analysis tasks after receiving training for both. Our experiments also provide evidence that 3D displays can better communicate spatial relationships, size, and shape of clusters.

Teaser: This figure contrasts the insights that can be gathered from a conventional point display (left) and our subspace shapes (right):

We show a high-dimensional point cloud projected into a generalized 3D subspace and visualized with a program called Subspace Voyager at an orientation selected using a virtual trackball. The points have been divided into five clusters tagged by color. Panel (a) shows a simple scatterplot projection; Panel (b) shows the same projection but with opacity depth cueing – points further back are rendered more transparent; Panel (c) shows our shaded display after converting the point cloud clusters into a set of solid models using the framework described in this paper.

Video: Watch it to experience how a simple point display gets transformed into our shaded subspace shapes

Paper:B. Wang and K. Mueller, "Subspace Shapes: Enhancing High-Dimensional Subspace Structures via Ambient Occlusion Shading," Extended abstract, IEEE Visualization, Berlin, Germany, October, 2018. pdf (extended version of the paper)