Pairwise Axis Ranking for Parallel Coordinates of Large Multivariate Data

Parallel coordinates has proven to be a scalable navigation framework for multivariate data. In parallel coordinate plots (PCP), human perception leverages spatial locality to determine the existence of multivariate patterns. A PCP can only show a handful of axes on most screens without cognitively overloading the user or obscuring patterns due to visual clutter. The traditional approach for axis ordering is to rely on the user to drag axes into positions to discover and elucidate a desired pattern. However, when data with thousands of variables are at hand, we do not have a comprehensive solution to algorithmically select the proper set of variables and order them to best uncover important or potentially insightful patterns. To further complicate the matter, important patterns may be dependent upon domain-specific properties or the scientific question at hand.

In this work, we developed a set of algorithms to rank axes based upon the importance of bivariate relationships among the variables as defined by a user-customizable metric. We provide an embarassingly parallel algorithm for computing the globally optimal ordering as well as a faster near-optimal algorithm. We showcase the efficacy of the proposed system by demonstrating autonomous detection of patterns. We demonstrate our approach using a new depth-enhanced rendering technique for a modern large-scale dataset of time-varying climate simulation. This work is currently under review for EuroVis09.

In the screenshot above, inverse correlation with consistent time constraints was used to automatically generate a PCP layout which relates the variance of radiation intensity on leaves as a function of the earth’s tilt throughout the seasons.

Credits:
The IPCC climate data was processed and analyzed by Forrest Hoffman, George Ostrouchov et al. under the auspices of Oak Ridge National Lab's Computer Science and Mathematics Division.