The analysis and prediction of eye gaze when viewing statistical graphs

Abstract

Statistical graphs are images that display quantitative information in a visual format that allows for the easy and consistent interpretation of the information. Often, statistical graphs are in the form of line graphs or bar graphs. In fields, such as cybersecurity, sets of statistical graphs are used to present complex information; however, the interpretation of these more complex graphs is often not obvious. Unless the viewer has been trained to understand each graph used, the interpretation of the data may be limited or incomplete [1]. In order to study the perception of statistical graphs, we tracked users’ eyes while studying simple statistical graphs. Participants studied a graph, and later viewed a graph purporting to be a subset of the data. They were asked to look for a substantive change in the meaning of the second graph compared to the first. To model where the participants would direct their attention, we ran several visual saliency models over the graphs [2–4]. Visual saliency models try to predict where people will look in an image; however, visual saliency models are typically designed and evaluated to predict where people look in natural images (images of natural or real world scenes), which have lots of potential information, subjective interpretations, and are not typically very quantitative. The ideal observer model [2], unlike most saliency models, tries to predict where people look based on the amount of information contained within each location in an image. The underlying theory of the ideal observer model is that when a person sees a new image, they want to understand that image as quickly as possible. To do this, the observer directs their attention first to the locations in the image that will provide the most information (i.e. give the best understanding of the information). Within this paper, we have analyzed the eye gaze from a study on statistical graphs to evaluate the consistency between participants in the way they gazed at graphs and how well a saliency model can predict where those people are likely to look in the graph. During the study, as a form of mental diversion to the primary task, participants also looked at natural images, between each set of graphs. When the participants looked at the images, they did so without guidance, i.e. they weren’t told to look at the images for any particular reason or objective. This allowed the viewing pattern for graphs to be compared to eye gaze data for the natural images, while also showing the differences, in the processing of simple graphs versus complex natural images. An interesting result shows that viewers processed the graphs differently than natural images. The center of the graph was not a strong predictor of attention. In natural images, a Gaussian kernel at the center of an image can achieve a receiver operating characteristic (ROC) score of over 80% due to inherent center bias in both the selection of natural images and the gaze patterns of participants [5]. This viewing pattern was present when participants looked at the natural images during the diversion task, but it was not present when they studied the graphs. Results from the study also found fairly consistent, but unusually low inter-subject consistency ROC scores. Inter-subject consistency is the ability to predict one participant’s gaze locations using the gaze positions of the other (n − 1) participants [3]. The saliency model itself was an inconsistent predictor of participants’ eye gaze by default. Like the participants, the saliency model identified titles and axis labels as salient. The saliency model also found the bars and lines on the graphs to be salient; however, the eye gaze of most participants rarely fell or focused on the line or bar graphs. This may be due to the simplicity of the graphs, implying that very little time or attention needed to be directed to the actual bar or line graph in order to remember it.