Visual encoding quality and scalability in information visualization

Abstract

Information visualization seeks to amplify cognition through interactive visual representations of data. It comprises human processes, such as perception and cognition, and computer processes, such as visual encoding. Visual encoding consists in mapping data variables to visual variables, and its quality is critical to the effectiveness of information visualizations. The scalability of a visual encoding is the extent to which its quality is preserved as the parameters of the data grow. Scalable encodings offer good support for basic analytical tasks at scale by carrying design decisions that consider the limits of human perception and cognition. In this thesis, I present three case studies that explore different aspects of visual encoding quality and scalability: information loss, perceptual scalability, and discriminability. In the first study, I leverage information theory to model encoding quality in terms of information content and complexity. I examine how information loss and clutter affect the scalability of hierarchical visualizations and contribute an information-theoretic algorithm for adjusting these factors in visualizations of large datasets. The second study centers on the question of whether a data property (outlierness) can be lost in the visual encoding process due to saliency interference with other visual variables. I designed a controlled experiment to measure the effectiveness of motion outlier detection in complex multivariate scatterplots. The results suggest a saliency deficit effect whereby global saliency undermines support to tasks that rely on local saliency. Finally, I investigate how discriminability, a classic visualization criterion, can explain recent empirical results on encoding effectiveness and provide the foundation for automated evaluation of visual encodings. I propose an approach for discriminability evaluation based on a perceptually motivated image similarity measure.