PROBLEM SPACE

The primary issues that are addressed are:
• scalability
• dimensionality
• “normalization”

Scalability

No matter how you slice and dice it, visualizing twin hierarchies of 200,000 nodes on a 14-inch monitor is no Polka party. When solving a problem of this magnitude, even the most accomplished InfoVis practitioner cannot help but wind up with occlusion, distortion, perceptual tricks, or an incomprehensible jumble of dots. There just aren’t enough pixels on the panel to show the data as is.

Is there a best way to promote examination of two large hierarchies and discern difference? Can it be done with a single window that provides both focus and context?

Context could be shown and details examined with a fisheye technique [11] or hyperbolic browser [8]. However, these techniques shift the position and rendering of surrounding nodes as the user changes focus. Due to the sheer size and complexity of these trees, it is desirable to maintain an overview that retains fixed context as the user interacts with the data. Zoomology utilizes separate windows for overview and detail. The two classification trees are merged into a single overview in which the hierarchy unfolds top to bottom and fans out alphabetically left to right. Color represents rank, and the width of a node is proportional to its number of descendants.

Due to the depth of each tree, it is still easy to lose track of one’s current position while viewing detail. Therefore, dual paths are provided to maintain context, both in the overview and in the legend.

Dimensionality

These classification trees contain three variables. One is rank, which is encoded by color. The other two are the Latin and common names, both of which are displayed as labels, as a mouseover tooltip in the overview and when the data reaches focus in the detail view.

Normalization

Because rank is crucial to interpretation of this dataset, we considered normalizing the tree along levels of rank for the overview. This model would not necessarily depict parents and children as residing at adjoining levels but would instead relegate all phylums to the same level, followed by all subphylums, then all superclass nodes, all class nodes, etc.

The programmers on our team favored a strict representation of the tree structure, while the non-programmers preferred a normalized version that emphasized the display of rank. Our most promising effort at normalization utilized concentric circles to represent the differing ranks. While this made it easy to distinguish rank, we could not find a clear method to represent a parent-child relationship that “skipped” a level. Our proposed solution was to leave blank space where a parent-child relationship skipped levels, but this did not seem intuitive and wasted space by necessitating empty areas in the visualization.

By distorting the tree structure, a normalized version also did not permit the user to answer some questions about the structure of the tree, such as “What is the deepest branch?”