Abstract
With the increasing availability and importance of data, the human visual system serves as a critical tool for analysis of patterns, trends, and relations among those data. Building on recent translational visual search work in domains like baggage screening (e.g., Mitroff et al., 2015) and radiology (e.g., Wolfe, 2016), we explored how different ways of representing data values can lead to efficient or inefficient visual processing of the relations between the values in a data pair. We asked participants to find a particular relation among the opposite relation (e.g., small/large value pairs among large/small), under a variety of common, and manipulated, data encoding methods: bullet graphs (one value as a bar, the other as a threshold dash through the bar), line graphs, connected dash graphs, area graphs, dash graphs (lines placed where the tops of bars in a bar graphs 'would be'), adjacent bars, and separated bars. Displays were divided into quadrants containing 1-5 data pairs each, for a total set size of 4-20 items. Participants were asked to quickly indicate which quadrant contained the target data pair. The choice of data depiction led to enormous differences in processing efficiency for relations between values (ranked fast to slow in the order listed above), from flat search slopes (line graphs), medium search slopes of 69 ms/pair (dash graphs), and severely steep search slopes of 115 ms/pair (separated bar graphs), one of the most ubiquitous encoding types. Visual search for relations can be strikingly serial, but performance can improve substantially with small changes to displays. Exploring visual search and relation processing in the context of data visualization displays may provide a rich case study for both basic research on the mechanisms of search, but also concrete guidelines for the students and scientists who use vision to process and convey patterns in data.
Meeting abstract presented at VSS 2017