Visualization Space Exploration : Theoretical and Practical Viewpoints

Abstract

Visualizations are graphical representations of data that have been used in a wide-ranging field of applications to provide a quick overview over data-inherent information. By taking advantage of human perceptual capabilities, visualizations help users understand features and phenomena in data, gather meaningful insights, and drive decision making processes. One of the main motivations in visualization research is to find the best visual representation for a given dataset, user, and task. This challenge is often solved in a subjective manner, where a visualization designer chooses graphical representations, visual channels, and encodings that they believe are best suited for the task at hand. Therefore, the effectiveness and reliability of the result largely varies with the designer's expertise. To make an objectively good design decision, the designer needs to consider all possible visualization methods, or in our words: explore the visualization space. For that purpose, the advantages and disadvantages of individual techniques can be highlighted through comparison methods based on quality metrics, user studies, or theoretical models. Each of these methods can additionally target the visual perception of representations, task-oriented and application-specific measures, structure-oriented matters, or meta-perceptual processes.

In this thesis, we aim to establish a greater understanding of the interconnections between independently studied approaches for visualization evaluation by exploring the visualization space from several different viewpoints. First, we take a theoretical approach to identify and classify previous work on the evaluation of visualization methods. We analyze theoretical models, user studies, and quality metrics, and combine them in a unified structure to distinguish classes of task-oriented, perceptual, meta-perceptual, and structure-oriented measures. We then describe the individual class strengths and shortcomings and propose a direction to combine the separate efforts into a bigger picture to advance the field of visualization research as a whole.

One instance where visualization exploration takes place in practice is during the development of visualization algorithms. By writing code, adding features, and changing parameters, visualization developers expose a large number of representations in visualization space. We developed a system that explicitly displays these individual states to the user and allows for their exploration and comparison. Parameter changes and their effect on all developed visualization states can be inspected to investigate their impact on visual features. The system not only encourages visualization developers to consider multiple representations when creating a visualization, but further allows for comparisons on a more general level. The simultaneous display of source code changes and visual changes in a meta visualization opens up a large branch of possible future research. We made a first step towards a practical development environment that encourages visualization comparison during the development process and reasoning about correlations of source code changes and their impact on the visual result.

In our implementation, we display source code states via node-link diagrams of their abstract syntax trees. Although this representation provides a clear outline of individual hierarchical structures, its juxtaposed nature impairs the comparison of many states. To overcome this issue, we analyzed existing methods for the visualization of dynamic hierarchies and combined the benefits of treemaps and stream-based approaches to display both the individual hierarchies and their evolution over time. We conducted a user study to evaluate the differences in effectiveness on low-level tasks and captured perceptual characteristics in hierarchical visualizations over time. The results suggest that our visualization can be applied as a general-purpose method to replace previous representations for static hierarchies and hierarchical changes over time. All compared visualization types and the effects of mutual parameters can be explored through our open-source implementation.

Finally, we explored aesthetic characteristics of artistic diagrammatic paintings and aimed to apply their visual appeal to storyline visualizations. We developed an interactive application that utilizes techniques for automatic layouting and image processing to create visual results similar to hand-drawn diagrams. Our application can further help artists create an initial layout by interactively adding data to the representation and focus their efforts on artistic aspects that are difficult for machines to imitate.

In the combination of our work, we explore the visualization field from several different viewpoints, move from visualization theory to practice, and show how individual components of visualization comparison can be combined for greater knowledge gain. We hope to encourage visualization researchers to merge their efforts into a larger theory and understanding of how visualizations work and to create objectively effective visualization solutions.