Well-designed visualizations leverage the human visual system to help people understand large data sets. Yet, producing effective visualizations is a challenging design task. Designers must carefully choose how to map the data to visual variables such as position, size, shape and color. In this process they make hundreds of nuanced judgments while balancing perceptual and cognitive tradeoffs. In response, researchers in psychology, cartography, statistics, and computer science have investigated the effects of different visual variables on graphical perception: the ability of viewers to interpret visual encodings and thereby decode information in graphs. Despite great progress in developing design guidelines based on laboratory experiments, comprehensive evaluation of the visualization design space and real-world validation of the resulting guidelines have remained elusive.
The research is advancing our understanding of graphical perception and formulate new guidelines for visualization design. The research involves new experiments to address unresolved issues in graphical perception, including large-scale web-based studies using crowdsourcing techniques and controlled laboratory studies using sensitive measurements, namely eye-tracking. The investigators are applying the results of these studies to (a) develop guidelines for effective visualization design, (b) instantiate these guidelines in automated design procedures, and (c) validate the guidelines and resulting tools through case study deployments.
This project has produced new experimental results that deepen our understanding of visualization design theory and provide actionable guidelines for design. We have also produced new systems and algorithms that support automated design and new ways to reuse existing designs. This work has led to a number of papers, some of which are already highly-cited. Four of the papers supported by this grant have been recognized as best papers or best paper nominees at the top conferences in Information Visualization and Human-Computer Interaction. Moreover, this grant has supported the development of multiple PhD and undergraduate students, as well as the release of open-source software tools, data sets, and teaching resources available for public use. Over the 2010-2011 academic year we focused on evaluating new methods for visualization evaluation (i.e., crowdsourced experimentation) and applying those methods to develop new findings and corresponding design guidelines. By replicating prior experiments, our results show that crowdsourcing platforms such as Amazon's Mechanical Turk enable inexpensive, largescale experimentation with valid results. We also developed additional guidelines to ensure high quality responses. We are now applying these methods in new experimental designs. For example, we have identified new results for improving the perceptual effectiveness of treemap displays, in the process overturning a decadesold assumption regarding the perceptual optimality of comparing square shapes. Our research has also taken first steps beyond the paradigm of lowlevel graphical perception: we have initiated an empirical investigation of narrative aspects of visualization design, uncovering both visual and interactive mechanisms of story telling with data. In addition we continue to design and develop graduate level courses on visualization and engage in public speaking about our results. In the 2011-2012 academic year we continued to advance our work on graphical perception focusing on the problem analyzing existing visualizations based on our knowledge of graphical perception. More specifically we worked on analyzing bitmap images of visualizations to extract the underlying marks and data from them. This analysis applies computer vision techniques based on our understanding of graphical perception. Once we have extracted the marks and data our system can automatically redesign the visualization to further facilitate graphical perception of the underlying information. We are also starting to conduct human-subject studies of scatterplots to identify the underlying principles for improving the perception of aggregate statistics such as mean, standard deviation and number of points. In the 2012-2013 academic year we developed more advanced techniques for using our understanding of graphical perception to analyze existing visualizations in the context of reports and webpages. Building on our previous ReVision system -- which could extract the underlying marks and data from an input bitmap of a chart -- we have developed a tool for overlaying the bitmap with additional information and metadata. In addition we have analyzed the space of such interactive chart overlays and developed a taxonomy for categorizing the different approaches. We have also recognized that most charts do not appear in a vacuum, but are embedded within text reports. Based on this insight we have been building a crowdsourcing approach for extracting the references between the marks in the chart and the surrounding text. We also extended our work on graphical perception to models and automated design methods for color selection. We have developed probabilistic models of color name-association from large crowdsourced data sets and shown how these models can be used to evaluate color design decisions in visualization. In subsequent work, we analyze the color distributions in results from from Google Image queries and use them to automate the choice of "semantically-resonant" colors which have better perceived correspondence with the underlying data. In a series of experiments we demonstrate that semantically-resonant assignment improves chart reading performance and that our algorithm is competitive with expert color designers.
