PhD Defense: Revealing Perceptual Proxies in Comparative Data Visualization

Talk
Brian Ondov
Time: 
12.15.2020 13:00 to 15:00
Location: 

Remote

Vision science has informed decades of development in the field of data visualization, providing theoretical and empirical bases for data to be translated into graphics in the most effective ways. Much of this work has focused on the relative efficacies of basic encodings, such as length, position, or area, in conveying quantities. Less is known, however, about what may affect comparison of multiple data series, which generally involves extraction of higher-order values, such as means, ranges, and correlations. In this work, we investigate such factors and the underlying visual processes that may account for them.We first investigate the influence of “arrangement”—that is, whether charts are shown side-by-side, stacked vertically, overlaid, etc.—on comparative tasks. In a series of psychophysical experiments, we vary task difficulty so participants achieve a target accuracy, allowing us to quantitatively compare efficacy of arrangements across tasks. These experiments span three chart types (bar, line, and donut), five arrangements (adjacent, stacked, mirrored, overlaid, and animated), and four discrimination tasks (item delta, correlation, mean, and range). The results suggest a complex interaction of factors, with different comparative arrangements providing benefits for different combinations of tasks and encodings. These observations in themselves offer some direct guidance to designers, and the methods provide a blueprint for investigating further conditions. We conclude, however, that the number of such interactions makes it infeasible to provide broadly applicable rankings of comparative arrangements, as has been done previously for encodings.Our subsequent efforts thus work instead toward understanding the visual processes that underly the extraction of statistical summaries needed for comparison. It is theorized that simpler shortcuts, called perceptual proxies, are used by the visual system to estimate these more complex values. For example, instead of computing the true means of two data series in a discrimination task, we may choose the series with the longest bar or with the largest convex hull area. We propose a list of candidate proxies and perform retrospective analyses, using data from the previous experiments, to narrow down a representative set for evaluation. We test the proxies with human subjects using an “adversarial” framework, in which the ranking of two charts along a task metric (e.g. mean) is opposite their ranking along a proxy metric (e.g. convex hull area). Perhaps chief among the results is the finding that use of a “centroid” proxy, among the ones tested, best explains the discrimination of means in bar charts.Finally, we attempt to characterize underlying visual processes of summary value estimation without any a priori assumptions (i.e. specific proxies), by using human-guided optimization to construct charts de novo. Exploring the multidimensional space of individual bar lengths, we use derivative-free, or “black box,” numerical methods to attempt to maximize the appearance of a summary statistic (here mean or range) to the human visual system, regardless of the true value. Quantitative interpretation of the resulting charts offers some corroboration of the centroid explanation for the mean, while qualitative interpretation may suggest additional proxies for range.This work, viewed as a whole, has both theoretical and practical implications. On the theoretical side, it contributes to perceptual psychology by offering evidence for underlying visual processes that may be involved in the interpretation of comparative visualizations. On the practical side, our experiments on arrangements provide straightforward guidance for displaying certain combinations of tasks and encodings, while our work with proxies is a step toward automated assessment of the deceptiveness of charts in comparative settings.
Examining Committee:

Chair: Dr. Niklas Elmqvist Dean's rep: Dr. Rob Patro Members: Dr. Leilani Battle

Dr. Eun Kyoung Choe Dr. John Dickerson Dr. Zhicheng Liu Dr. Adam Phillippy