next up previous
Next: Future Work Up: H3: Laying Out Large Previous: Implementation

Analysis

 

The H3 layout technique can easily handle thousands of nodes and has been tested on graphs of over 20,000 nodes. It is very effective at presenting a large neighborhood of a huge graph in a small amount of screen space. Plate 1 gif shows a medium sized Web site which contains 5000 total nodes. The leftmost image shows a selected node with outgoing nontree links drawn. Although distant subtrees are quite distorted, we can see enough context that the destinations of the non-tree links can be roughly distinguished. In the other two images we bring the cluster of nodes which contains the destination of most of the nontree links closer to the focus. Notice that we can still see the originating node, although it is quite far away in the tree structure.

Static pictures such as figures in this paper may be misinterpreted as showing 2D objects on the surface of a hemisphere instead of as 3D objects inside the volume of a ball. In the interactive system the scene is disambiguated as soon as the user sees the objects rotate inside the ball. The interactive user experience is difficult to communicate through still images, but can be approximated with image sequences and the accompanying video.

Hyperbolic methods are very effective at providing global overviews and displaying many nodes at once. We can categorize the drawn nodes into three classes: main/labelled, peripherial, and fringe. What we call peripheral nodes are small but still distinguishable as individual entities upon close inspection. Fringe nodes are not individually distinguishable, but their aggregate presence or absence shows significant structure of far-away parts of the graph. Each class can fit roughly an order of magnitude more than the last. The H3 and PARC browsers can both show up to 50 main/labelled nodes. The PARC layout doesn't have peripheral nodes as such, since nodes are not drawn as discrete entities. The H3 layout can show up to 500 peripheral nodes. The H3 fringe can show information about thousands of nodes, whereas the PARC fringe shows information about hundreds of nodes. (The webviz browser [MB95] didn't have labelled nodes, could handle up to 50 peripheral nodes and show information about up to 500 fringe nodes.)

Another big advantage of moving from a 2D tree to a 3D graph is the ability to see non-tree links in context. One of the greatest strengths of the H3 approach is the ability to see relationships between a part and the far-flung reaches of the whole. Although the details of the nontree link destinations are usually distorted, a rough sense of their direction helps the user construct and maintain a mental model of the overall graph structure. The details become clear in a smooth transition when that area of the structure is brought towards the center. It might be possible to extend a 2D browser to support graph display, but nontree links would have to be drawn in the same plane as the main spanning tree links, necessarily intersecting them. In the 3D system the nontree links can follow paths which are unlikely to intersect the surrounding spanning tree links.

Information density is not the only metric: when taken too far, it becomes clutter. Going to the third dimension is not a panacea. Drawing all the links in a highly connected 3D graph yields a picture which can give a high level overview of the global structure but is useless for examining the details. The ability to interactively select which non-tree links are drawn is a major improvement of the SiteMgr system over the webviz system. Plate 2 gif shows a function call graph for a scientific computing benchmark. Such graphs are notoriously difficult to understand when all the links are drawn. Software engineers who must modify or optimize unfamiliar code can browse through the H3 layout of a call graph to understand structure of a complex program.

The H3 system is very effective at showing aspects of a large graph such as overviews and part/whole relationships that are poorly displayed using other techniques. The converse is that H3 is not well suited for tasks where traditional systems shine, like selecting an item from a linear list. We promote the H3 layout as an additional module to augment other visualization components, not a general-purpose browser that should replace all other views. The SiteMgr system also includes a traditional 2D directory browser to provide an different view. Nodes selected in one browser are automatically highlighted (brushed) in the other, and each can trigger scrolling or an animated transition to the selected document in the other.


next up previous
Next: Future Work Up: H3: Laying Out Large Previous: Implementation

Tamara Munzner Tue Jul 22 08:42:41 PDT 1997