MAVIS is a (soon to be released) public-domain software package designed by Chris Bentley and Matthew Ward at WPI for the interactive visual exploration of multivariate data sets. Multidimensional Scaling (MDS) is used to reduce data of arbitrary dimensionality to 1, 2, or 3 dimensions for display. MAVIS animates the optimization process to allow viewers to see interpoint relationships (stress) and overall quality of the positioning of points. The software is portable to all UNIX platforms which support XR4 or higher. The current version of the software (1.0) supports the following functionality:
MAVIS will soon be available via anonymous ftp at cs.wpi.edu in the directory pub/projects_and_papers/graphics_and_vision/visualization.
An overview of MAVIS was presented at the 1996 IEEE Information Visualization Symposium (abstract) .
The slides from Chris Bentley's thesis presentation are also available.
Here is an animation of the Iris data set. Note that even after things have settled down, there is still some stress (red, purple) in the system.
Here is another animation, consisting of 20 117-dimensional points representing benchmark results (courtesy of Jozo Dujmovic, San Francisco State University). Note the outlier (col 12) as well as several pairs that maintain close proximity.
Some screen dumps of MAVIS in operation are provided below. The data used is either the IRIS data (5 dimensions, 150 data points) or a set of benchmarking results (120 dimensions, 20 data points).
In this figure, the bonds between nodes are displayed, colored based on the level of stress along the bond (blue is low, red is high)
Here we show the tracks of motion (early in animation) to look for nodes which are moving in a similar fashion.
In this figure, we don't clear the screen after each iteration. This gives another view of the history of motion.
We can also display the labels of nodes (if available), although it gets kind of messy with large data sets unless you zoom in on a subsection of the display.
MAVIS currently provides three additional plots of the animation process. The first is a Sheppard plot, which shows 2-D distances plotted against N-D distances for the current frame of the animation.
The Stress plot shows the average stress at each node over the life of the simulation. Ideally this should be monotonically decreasing, but since some motions which are favorable for some nodes are stressful to others, the plot can jump around a bit.
Finally, the Correlation plot shows how well the 2-D and N-D distances agree through the life of the animation.
Matthew O. Ward (email@example.com)