Investigating visualizations with psychophysics : understanding the perception of correlation in two-class scatterplots

Information visualization is an increasingly important approach to handling the endless stream of data in work and daily life. Visualizations have evolved for use by the human visual system, and they capitalize on our perceptual and cognitive abilities to help us see and understand patterns in data displays. Recently, vision science methods have been used to evaluate visualizations and provide empirical guidelines for their design. In this dissertation, we argue for a two-way linkage between vision scientists and visualization researchers. We demonstrate the value of this work for both communities in a multi-experiment case study about the perception of correlation in two-class scatterplots. Although we have a basic understanding of the perception of correlation in single-class scatterplots, multi-class scatterplots have not been examined to the same extent. Our work examined this issue by using recent color modelling techniques to investigate correlation perception in two-class scatterplots. In our studies, participants performed correlation discriminations on side-by-side scatterplots, each containing a “target” population to evaluate, and a “distractor” population to ignore. We also compared correlation perception with numerosity perception in two experiments. Results show that displaying a second class in a scatterplot incurs a robust performance cost in perceiving correlation, and that the nature of this performance cost is dependent on the task. Greater interference for target populations, or less precise discrimination performance, occurs for correlation perception versus numerosity perception in the same display. Furthermore, if the extent of the target class is greater than that of the distractor class, this interference disappears. We conclude that, contrary to intuition, perceptual performance does not depend on feature differences at all; our visual system simply cannot use color or shape information to select sub-populations in two-class scatterplots when estimating correlation.Arts, Faculty ofPsychology, Department ofGraduat