Small multiples are a popular visualization method, displaying different views of a dataset using multiple frames, often with the same scale and axes. However, there is a need to address their potential constraints, especially in the context of human cognitive capacity limits. These limits dictate the maximum information our mind can process at once. We explore the issue of capacity limitation by testing competing theories that describe how the number of frames shown in a display, the scale of the frames, and time constraints impact user performance with small multiples of line charts in an energy grid scenario. In two online studies (Experiment 1 n = 141 and Experiment 2 n = 360) and a follow-up eye-tracking analysis (n = 5), we found a linear decline in accuracy with increasing frames across seven tasks, which was not fully explained by differences in frame size, suggesting visual search challenges. Moreover, the studies demonstrate that highlighting specific frames can mitigate some visual search difficulties but, surprisingly, not eliminate them. This research offers insights into optimizing the utility of small multiples by aligning them with human limitations.

Many papers make claims about specific visualization techniques that are said to enhance or calibrate trust in AI systems. But a design choice that enhances trust in some cases appears to damage it in others. In this paper, we explore this inherent duality through an analogy with "knobs". Turning a knob too far in one direction may result in under-trust, too far in the other, over-trust or, turned up further still, in a confusing distortion. While the designs or so-called "knobs" are not inherently evil, they can be misused or used in an adversarial context and thereby manipulated to mislead users or promote unwarranted levels of trust in AI systems. When a visualization that has no meaningful connection with the underlying model or data is employed to enhance trust, we refer to the result as "trust junk." From a review of 65 papers, we identify nine commonly made claims about trust calibration. We synthesize them into a framework of knobs that can be used for good or "evil," and distill our findings into observed pitfalls for the responsible design of human-AI systems.

Abstract People frequently make decisions based on uncertain information. Prior research has shown that visualizations of uncertainty can help to support better decision making. However, research has also shown that different representations of the same information can lead to different patterns of decision making. It is crucial for researchers to develop a better scientific understanding of when, why and how different representations of uncertainty lead viewers to make different decisions. This paper seeks to address this need by comparing geospatial visualizations of wildfire risk to verbal descriptions of the same risk. In three experiments, we manipulated the specificity of the uncertain information as well as the visual cues used to encode risk in the visualizations. All three experiments found that participants were more likely to evacuate in response to a hypothetical wildfire if the risk information was presented verbally. When the risk was presented visually, participants were less likely to evacuate, particularly when transparency was used to encode the risk information. Experiment 1 showed that evacuation rates were lower for transparency maps than for other types of visualizations. Experiments 2 and 3 sought to replicate this effect and to test how it related to other factors. Experiment 2 varied the hue used for the transparency maps and Experiment 3 manipulated the salience of the borders between the different risk levels. These experiments showed lower evacuation rates in response to transparency maps regardless of hue. The effect was partially, but not entirely, mitigated by adding salient borders to the transparency maps. Taken together, these experiments show that using transparency to encode information about risk can lead to very different patterns of decision making than other encodings of the same information.

In theory, visual saliency in a graph should be used to draw attention to its most important component(s). Thus salience is commonly viewed both as a basis for predicting where graph readers are likely to look, and as a core design technique for emphasizing what a reader is intended to see among competing elements in a given chart or plot. We briefly review models, metrics, and applicable theories as they pertain to graphs. We then introduce new saliency models based on perceptual and cognitive theories that, to our knowledge, have not been previously applied to models for viewing statistical graphics. The resulting frameworks can be broadly classified as bottom-up perceptual models or top-down cognitive models. We report the results of evaluating these new theory-informed approaches on gaze data collected for statistical graphs and for more general information visualizations. Interestingly, the new models fare no better than previous ones. We review the experience, noting why we expected these hypotheses to be effective, and discuss how and why their performance did not match our aspirations. We suggest directions for future research that may help to clarify some of the issues raised.

To interpret data visualizations, people must determine how visual features map onto concepts. For example, to interpret colormaps, people must determine how dimensions of color (e.g., lightness, hue) map onto quantities of a given measure (e.g., brain activity, correlation magnitude). This process is easier when the encoded mappings in the visualization match people's predictions of how visual features will map onto concepts, their inferred mappings. To harness this principle in visualization design, it is necessary to understand what factors determine people's inferred mappings. In this study, we investigated how inferred color-quantity mappings for colormap data visualizations were influenced by the background color. Prior literature presents seemingly conflicting accounts of how the background color affects inferred color-quantity mappings. The present results help resolve those conflicts, demonstrating that sometimes the background has an effect and sometimes it does not, depending on whether the colormap appears to vary in opacity. When there is no apparent variation in opacity, participants infer that darker colors map to larger quantities (dark-is-more bias). As apparent variation in opacity increases, participants become biased toward inferring that more opaque colors map to larger quantities (opaque-is-more bias). These biases work together on light backgrounds and conflict on dark backgrounds. Under such conflicts, the opaque-is-more bias can negate, or even supersede the dark-is-more bias. The results suggest that if a design goal is to produce colormaps that match people's inferred mappings and are robust to changes in background color, it is beneficial to use colormaps that will not appear to vary in opacity on any background color, and to encode larger quantities in darker colors.

Evaluating the effectiveness of data visualizations is a challenging undertaking and often relies on one-off studies that test a visualization in the context of one specific task. Researchers across the fields of data science, visualization, and human-computer interaction are calling for foundational tools and principles that could be applied to assessing the effectiveness of data visualizations in a more rapid and generalizable manner. One possibility for such a tool is a model of visual saliency for data visualizations. Visual saliency models are typically based on the properties of the human visual cortex and predict which areas of a scene have visual features (e.g. color, luminance, edges) that are likely to draw a viewer's attention. While these models can accurately predict where viewers will look in a natural scene, they typically do not perform well for abstract data visualizations. In this paper, we discuss the reasons for the poor performance of existing saliency models when applied to data visualizations. We introduce the Data Visualization Saliency (DVS) model, a saliency model tailored to address some of these weaknesses, and we test the performance of the DVS model and existing saliency models by comparing the saliency maps produced by the models to eye tracking data obtained from human viewers. Finally, we describe how modified saliency models could be used as general tools for assessing the effectiveness of visualizations, including the strengths and weaknesses of this approach.

From the seminal work of Yarbus [1967] on the relationship of eye movements to vision, scanpath analysis has been recognized as a window into the mind. Computationally, characterizing the scanpath, the sequential and spatial dependencies between eye positions, has been demanding. We sought a method that could extract scanpath trajectory information from raw eye movement data without assumptions defining fixations and regions of interest. We adapted a set of libraries that perform multidimensional clustering on geometric features derived from large volumes of spatiotemporal data to eye movement data in an approach we call GazeAppraise. To validate the capabilities of GazeAppraise for scanpath analysis, we collected eye tracking data from 41 participants while they completed four smooth pursuit tracking tasks. Unsupervised cluster analysis on the features revealed that 162 of 164 recorded scanpaths were categorized into one of four clusters and the remaining two scanpaths were not categorized (recall/sensitivity=98.8%). All of the categorized scanpaths were grouped only with other scanpaths elicited by the same task (precision=100%). GazeAppraise offers a unique approach to the categorization of scanpaths that may be particularly useful in dynamic environments and in visual search tasks requiring systematic search strategies.

Abstract Effectively evaluating visualization techniques is a difficult task often assessed through feedback from user studies and expert evaluations. This work presents an alternative approach to visualization evaluation in which brain activity is passively recorded using electroencephalography (EEG). These measurements are used to compare different visualization techniques in terms of the burden they place on a viewer's cognitive resources. In this paper, EEG signals and response times are recorded while users interpret different representations of data distributions. This information is processed to provide insight into the cognitive load imposed on the viewer. This paper describes the design of the user study performed, the extraction of cognitive load measures from EEG data, and how those measures are used to quantitatively evaluate the effectiveness of visualizations.