History of XmdvTool

XmdvTool originated in 1993, when Matt Ward decided to merge several small programs he had written to visualize multivariate data in order to evaluate the effectiveness of each. The original techniques included were parallel coordinates, scatterplots, and star glyphs. He later added dimensional stacking, a technique he and two of his students (Jeff LeBlanc and Rajeev Tipnis) had created and implemented in a tool called N-Land. The techniques were linked via a highlighting mechanism controlled by N sliders, one for each dimension. Version 1.0 was developed using Athena Widgets and the
Widget Creation Library. It had about 1000 lines of C code and about 600 lines of a resource file to define the interface. It was released to the public domain during the summer of 1994, and described in a paper at the Visualization '94 conference.

One of the reviewers of this paper commented that the program really needed direct manipulation to support the selection/highlighting process. Thus Ward and Allen Martin, a WPI graduate student, added functionality to support interactive specification and modification of the selection region, called the "brush", on the data displays themselves, rather than via separate widgets. This work formalized the notion of an N-dimensional brush, as opposed to the screen-space techniques that had been used before. By working in data-space rather than screen-space, users could specify a hyperbox of interest, which could then be highlighted, deleted, or masked. Also, the concepts of a ramped brush, which allowed points to belong to a brush in varying degrees, and a composite brush, built from logical combinations of existing brushes, weredeveloped. This work was reported at Visualization '95, and version 2.0 of the system was released that year. The code had grown to approximately 10,000 lines of C code and 1000 lines of resource file.

Because the source code was distributed free-of-charge, several researchers made their own variations on the code. Torsten Timm, a student at the University of Erlangen, released the unofficial version 3.0 in 1996. His extensions allowed data-driven glyph positioning, labeling of data points, and other enhancements.

Version 3.1, developed at WPI, included many features beyond those in Version 2.0, including interactive enabling and disabling of dimensions, arbitrary zooming, extended Tukey box plots for visualizing statistics of selected data, a reconfigured user interface, and an improved help facility. It was released in 1997.

In 1998, Matt Ward and Elke Rundensteiner received a 3-year NSF grant to investigate ways of visually exploring large multivariate datasets. The award, NSF:IIS 9732897, is being used to fund 2 graduate students: one in visualization and the other in databases. The key concept of the research is to use multiresolution techniques and hierarchical clustering/partitioning methods to make large dataset analysis more manageable.

Version 4.0, developed by Ying-Huey Fua at WPI, represented a major change from numerous perspectives, not the least of which was that it no longer depended on the X-Window environment. Using OpenGL and Tcl/tk, the package would now run on both Unix and Windows platforms. A new interface, featuring Windows-like menus and icons to represent the different display modes, has a very different look-and-feel from previous versions. This version also introduced a new method for exploring large multivariate data sets, based on hierarchically clustering the data and presenting aggregation information rather than just the raw data. Hierarchical parallel coordinates, described in a paper at Visualization '99, show cluster summarizations via shaded bands rather than single polylines, conveying both cluster extents and size. Bands are colored via their location in the hierarchy, such that child and parent clusters have similar clusters. A new brushing paradigm, called structure-based brushing, was also introduced. This allows users to navigate the hierarchy, selecting subtrees for which they would like to display more or less detail. This new technique for interaction was reported in a paper at Information Visualization '99, and also in an extended paper in IEEE Transactions on Visualization and Computer Graphics.

The most recent version, 4.1 (to be released soon), extends the support for hierarchically structured data to scatterplot matrices, star glyphs, and dimensional stacking. Using the same coloring and navigation strategies pioneered in Version 4.0, Jing Yang has created a consistent and complete strategy for large-scale multivariate data exploration. She has also
eradicated several bugs found in version 4.0, such as bands extending beyond the drawing area boundaries, and made many improvements in the efficiency of the code.

Version 4.2, to be released during the Fall of 2000, will finally tie together the visualization front-end with a database backend. Using a novel indexing structure for hierarchies, Daniel Stroe has developed efficient query mechanisms for all navigation operations on the structure-based brush. He has also experimented with caching and prefetching techniques to further improve performance. All of these additions have been integrated into an Oracle backend which communicates seamlessly with the existing visualization front-end.

Beyond that, who knows? We have lots of ideas for future directions, including ways to deal with large numbers of dimensions, user-guided reclustering, non-numeric data, and a host of other suggestions made by current users. We hope to have a major release every 18 months or so, with minor releases as warranted.

Acknowledgement

The XmdvTool Project is supported by NSF under grant IIS-9732897, the NSF CISE Instrumentation grant IRIS 97-29878 and the NSF grant IIS-0119276.