There are various styles to visually represent relational data in the form of node-link diagrams. In particular, for planar graphs we can find orthogonal node-link diagrams consisting of links bending only at ninety degrees a successful and prominent variant. One of the benefits of such drawings is the tracking of longer paths through a network with the eyes due to their limited number of link orientations, changes, and variations, but on the negative side the links can have arbitrary bending shapes. In this article we developed a novel way to visualize such orthogonal planar drawings by making use of mazes that look more natural to the human eye due to the street-like visual metaphor that many people are familiar with. Tracking paths is one of the major tasks in such graph visualizations, similar to orthogonal node-link diagrams, however, we argue that mazes are a more natural way to find paths. To get insights in the visual scanning behavior when reading graph mazes we conducted a comparative eye tracking study with 26 male versus female participants of different experience levels while also alternating between orthogonal node-link drawings and graph mazes as well as different graph size levels. The major result of this comparative study is that the participants can track paths in both representation styles, including a geodesic path tendency in their visual search behavior, but typically have a longer fixation duration at branching nodes and locations in the mazes that lead to opposite directions to the geodesic path tendency, maybe the viewers had to start a reorientation phase in their visual scanning behavior. We also found out that the size, that is the number of graph vertices has an impact on the visual scanning behavior for both orthogonal node-link diagrams as well as street-like maze representations, but for the mazes we found this impact to be less strong (in terms of the eye movement data metrics fixation durations and saccade lengths) compared to the node-link diagrams. To conclude the article, we discuss limitations and scalability issues of our approach. Moreover, we give an outlook and future work for possible extensions.
Read full abstract