Thesis Defense


"A Computational Approach to Mitigate Visualization Design Barriers"

Connor Gramazio

Friday, April 21, 2017 at 10:30 A.M.

Lubrano Conference Room (CIT 4th Floor)

I will present visualization techniques that have improved cancer genomics applications by automatically (1) analyzing task requirements and (2) creating effective color palettes. These offer evidence that computational techniques can make usable design practices more accessible to visualization creators.

I will first present an evaluation of how simple classifiers trained on annotated mouse interaction logs can help designers understand how domain experts use visualizations. This discussion will also briefly delve into how such approaches can complement traditional qualitative user studies. I will also provide evidence that classification approaches, when paired with traditional qualitative methods, can help mitigate perennial domain expert evaluation concerns such as sample and observation bias, while also validating and expanding the generalizability of observational findings.

In the second half of the talk, I will shift focus to how graphical perception theory can inform computational design. Specifically, I will overview an evaluation of Colorgorical: a tool for creating categorical information visualization color palettes. I will discuss evaluation results showing that Colorgorical can automatically produce palettes that are as discriminable and typically more preferable compared to those included in ColorBrewer, Microsoft, and Tableau. I will also explain how Colorgorical achieves these results by applying vision science theory in an iterative semi-random sampling procedure to predict color combinations that satisfy user-defined balances of discriminability and preference.

These contributions reduce usable design adoption barriers by mitigating the expertise and labor typically required for user-centered design (e.g., the logistical difficulty of recruiting domain experts for study). Finally, given the ubiquity of these usability issues across domain applications, these contributions also have potential to inform visualization design beyond cancer genomics applications, such as in the brain sciences.

Host: David Laidlaw