A Motion to Consider: New Evidence for the Value of Animation in Visual Analysis

By Brian Ondov and Niklas Elmqvist

Animation in data visualization has been through it all, from initial hype, to explosion of applications, to skepticism and debunking. If nothing else, though, the topic of animation is persistent, perennially asserting its presence both in design and in the scientific community. In fact, a landmark study casting doubt on the efficacy of animation will receive a 10-year Test of Time award at the upcoming IEEE VIS conference, emphasizing the enduring relevance of this mechanism, even in the face of negative results.

Motion seems speaks to us at a fundamental level, perhaps because it is among the most primitive and fundamental elements of vision. Perception of motion originates in the retina itself, and even motion that is outside our visual field can trigger an innate response to look toward it. But can animation have perceptual value beyond just catching our attention?

As is often the case in science, we stumbled upon an interesting result almost by accident. Our intent was to investigate the usefulness of different layouts to support comparison, with an especially curious eye toward symmetric arrangements, such as population pyramids. For the sake of completeness, we threw in animation as a bit of a wild card. To our surprise, in our first experiment, animation won the day, enabling participants to perform the task better even than with superimposed displays (which were intended as a control!).

So what is going on here? Are these results incongruent with other experiments? Not necessarily. While it is tempting to use perceptual studies to label visual techniques as “good” or “bad”, the reality is of course more complicated. For example, our setup involved only a small number of moving shapes, and required the ranking of very subtle differences between data points. Both of these factors likely made it easier to track and compare elements of the scene. We think this allowed participants to bring a powerful perceptual ability to bear: the estimation of how fast an object is moving.

What does this mean going forward? Well, for one, our study was highly controlled and narrowly focused, and there is plenty more to be done to explore all the factors that may be at play. More broadly, though, we may benefit simply from adopting a more nuanced view the value of animation for conveying information. One thing that is clear is the continued importance of answering these questions as the field progresses. Whether it’s used to enhance cognition or just for splash, animation probably isn’t going anywhere.

More information about this work:

Ondov, Brian, Nicole Jardine, Niklas Elmqvist, and Steven Franconeri. “Face to Face: Evaluating Visual Comparison.” IEEE Transactions on Visualization and Computer Graphics (2018).

To learn more, contact Brian Ondov at ondovb@umd.edu.

For video and demos, visit: https://hcil.umd.edu/visualcomparison/

To learn more about the HCIL, please visit: http://hcil.umd.edu/

A Moment of Reflection: How Symmetry Can Help Us Interpret Charts

By Brian Ondov and Niklas Elmqvist

Symmetry has a long history of study in perceptual psychology, from the Gestalt movement to modern, computerized experiments with flashing point clouds. Researchers have even tested whether this basic element of visual organization is still perceived by astronauts in microgravity (it is!). Much less studied, though, is how our innate ability to recognize symmetrical shapes and scenes might affect how we see data in visualizations.

Perhaps recognizing the power of symmetry intuitively, demographers have long used it to juxtapose the male and female components of population pyramids, beginning in the late 19th Century. Here at the HCIL, though, we arrived at symmetry from a somewhat different angle, while tackling the problem of how to compare two sunburst charts. Nonetheless, we saw an opportunity to provide more experimental evidence for the technique (and, in fact, for comparative displays more generally).

We asked two main questions: (1) does symmetry help pick out a “biggest mover” between two datasets (top, left), and (2) does it help identify overall similarity (top, right)?

The results were promising: for the first task, the symmetrical arrangement indeed allowed participants to identify more accurately which bar changed the most, compared to a typical side-by-side view. This supports the idea that we can see not only whether a shape is symmetrical, but also which parts are or are not. Symmetry, though, was not the top performer here—the task was even easier using a superimposed display, and easier still with animation instead of two static views (the latter was a pleasantly incidental finding).

Where symmetry really shined, though, was in the second task, involving overall similarity. Here, the symmetrical arrangement outperformed all others, including superimposing and animating. This is perhaps not so surprising, given what we know about perception. After all, when arranged in this way, more similar data sets will create more symmetrical shapes. Still, this provides some empirical evidence that had been missing, which is pretty exciting for those of us that traffic in data. On top of that, as you may have noticed, the charts in this task look a lot like the population pyramids mentioned earlier. This is a nicely mutual validation (a symmetry, if you will): it’s both experimental support for a long-used technique and practical corroboration of our controlled experiment.

Of course, these were very focused studies, with the number of variables intentionally limited. In the future, we can ask a host of other questions to tease out where, when, and why symmetry works—and doesn’t work—in data visualization. Naturally, the technique will not be appropriate for all situations, especially if more than two datasets need to be compared. If nothing else, these results have given us something to reflect on.

More information about this work:

Ondov, Brian, Nicole Jardine, Niklas Elmqvist, and Steven Franconeri. “Face to Face: Evaluating Visual Comparison.” IEEE Transactions on Visualization and Computer Graphics (2018).

To learn more, contact Brian Ondov at ondovb@umd.edu.

For video and demos, visit: https://hcil.umd.edu/visualcomparison/

To learn more about the HCIL, please visit: http://hcil.umd.edu/

 

Understanding Data by Having Computers and Humans Work Together

By Zhe Cui, Karthik Badam, Adil Yalcin, and Niklas Elmqvist

Understanding data using computer data science tools can be seen as a collaboration between the analyst and the computer. Most data tools put the analyst in the driver’s seat to guide the analysis. For example, in Excel, you need to select rows and columns and choose which chart type to use to calculate values, create charts, and develop insights. Such tools completely rely on the user to understand the data. However, this one-sided arrangement is not sufficient for modern large datasets and complex analytical scenarios when (1) the user is unaware of the best methods to transform, arrange, and analyze the data, or (2) is overwhelmed by the sheer scale and complexity of the data (or both).

In a recent research project, we explore an alternative approach that supports the analyst by providing automatic insights from the data through visualizations. This proactive approach leverages any available computer power to run data analyses automatically in the background and present insights (or “a-ha moments”) that can aid the analyst’s exploration. Our tool, DataSite, is an example of this idea that is designed to improve the coverage of the data the user views, the qualities of insight collected, as well as the user’s engagement during data analysis.

The Role of Automation in Generating Insights

An insight is an observation of value made from data in the sensemaking process. In visual sensemaking, people create data visualizations through tools such as Microsoft Excel and the industry-standard tool Tableau to extract trends, patterns, and outliers, eventually leading them to form insights. However, this process can be overwhelming if performed manually when dealing with a large number data items with many attributes, for example all of the students and their data enrolled at a university, or the sales data over time for a multinational company. The main challenge here is that the user will not know what the visualization looks like until it is shown, and this “trial and error” process can therefore be long and exhausting.

At the core of our work is the idea that computers can alleviate this burden by automatically generating observations and “useful” charts. After all, computing has always been about simplifying people’s lives. The same should be true of data analysis.

While this is thought-provoking, there is no perfect definition of an insight yet (John Stasko discussed this aspect more specifically in his recent blog post), which complicates our goal of proactive computation. However, “insight” has often been a subjective word and heavily depends on the goals of analysis and the user, as well as the domain. For instance, a paper from Tang et al. at SIGMOD 2017 developed algorithms to extract top insights from a dataset based on a database perspective in order to help enterprises make better and faster decisions.

Blending the Best of Human and Machine Capabilities

Our proactive approach is based on the core philosophy that “human thinking is expensive, whereas computational resources are cheap.” DataSite utilizes computer resources such that when the user analyzes and visualizes the data, the computer or server (or even a cluster) simultaneously executes appropriate automatic analyses on the data in the background to suggest interesting leads to the user to investigate as a next step. For instance, while observing the differences between cars with different horsepower, DataSite suggests differences in Miles per Gallon based on a correlation analysis that was automatically executed in the background.

 

By continuously executing all conceivable analyses on all combinations of data dimensions, DataSite uses brute force to generate automatic insights. This enhances the analyst’s awareness of the data in the data exploration process by both choosing best practice analysis methods as well as eliminating the need for the human to perform costly calculations. The computer-generated insights are presented through a user interface element called the feed view, which streams continuously updating results from computational analyses, akin to social media feed such as on Twitter or Facebook. Each result is accompanied by a visualization that highlights the result in the context of the data items. This leads to an analytical workflow that mixes the best of human and machine intelligence.

Better Insights and User Engagement through Proactive Insights

We evaluated our approach in DataSite through two user studies of open-ended visual exploration. In these studies, we compared DataSite to manual visualization (Polestar) and visualization recommendation (Voyager 2), respectively. In DataSite, we focused on standard automated analyses such as computation of statistical measures for mean, variance, and frequency of items, and standard data science methods for clustering, regression, correlation, and dimension reduction. The task for studies is the same: exploratory analysis of unknown data (also called “open-ended task”). We used 2 tools with 2 datasets (one dataset on each tool interface). Participants started with one tool and dataset, and then moved to the second interface. They were asked to explore the dataset “as much as possible” within a given time of 20 minutes and were encouraged to speak out aloud their thinking process and insights. Three major benefits of DataSite emerged from this study:

  • Broader coverage: DataSite shows 30% increase in data field attribute coverage compared with Polestar. There are more multi-attribute charts (encoding two or more data attributes) that participants viewed and interacted with using DataSite than Polestar. When compared with Voyager 2, DataSite has comparable data field attribute coverage but provided more meaningful charts.
  • More time spent on charts: Most participants spent at least 25% of their time on exploring the feed itself. All participants felt that the feed is useful for analysis and provides guidance of “where to look” in the data.
  • Better subjective ratings: People rated DataSite more efficient and comprehensive than Polestar and Voyager 2.

The Future of Proactive Analytics

 DataSite can be seen as a canonical visual analytics system in that it blends automatic computations with manual visual exploration. We regard it as the first step towards a fully proactive visualization system involving explicit human feedback in the loop, such as tasks people are doing, data attributes people care about, and advanced analysis people want to dive into. Besides, inferential statistics and user behavior based recommendations can also be integrated to provide user-guided recommendations of insights. A truly intelligent visual analysis system would leverage possible feedback from user and computational power from the computer to present easily understanding and interpretable insights.

More Resources

Paper: https://arxiv.org/abs/1802.08621

Advanced Transportation Lab is a long term partner of HCIL

At the University of Maryland’s Center for Advanced Transportation Technology Laboratory (CATT LAB), more than 100 engineers, developers, researchers and students work to provide officials with actionable insight that can help keep transportation systems running smoothly. Housed in the A. James Clark School of EngineeringThe CATT Lab is interested in providing planners, operators and researchers with computer user interfaces that leverage big data to create robust visualizations—charts, graphics and tables—that can assist in transportation decision-making.

Catherine Plaisant, Senior Research Scientist at the University of Maryland Institute for Advanced Computer Studies and Associate Director of Research of the Human-Computer Interaction Lab,  is part of the CATT Lab’s User Experience (UX) Team. The team is tasked with creating novel user interfaces for the Regional Integrated Transportation Information System.  From a major snowstorm to an unexpected event requiring multiple road closures or detours, disruptions to our nation’s highways and connecting transportation systems can often lead to big problems and major delays—frustrating not only travelers, but also law enforcement and transportation officials. The CATT lab can not only only pinpoint the worst bottlenecks, but identify what caused them, how long the backups are, and how much they cost, and make recommendations to the state of Maryland. “Research findings often take years or decades to impact the products people use, while the work with the CATT lab is rapidly implemented by a team of excellent developers and used within weeks by hundreds of operators and managers in transportation agencies across the U.S.,” Catherine says.

John Allen, a faculty assistant at CATT, calls Plaisant an “integral part” of the UX Team. “Catherine brings clear insight, intelligence and an innate talent to conceptualizing and developing interfaces and visualizations, critical to useful and usable next-gen analytics,” he says. In particular, Allen says, Plaisant’s contributions and guidance have made RITIS tools easy to understand, simple to use, with great actionable visualizations that help those in the industry quickly identify problems and develop smart, cost-effective mobility, safety and security solutions.  For instance, Plaisant has been involved in the creation of an Origin-Destination Data Suite, which will gain trip insight derived from leveraged geospatial data. This enables unprecedented understanding of vehicular movement information, such as origin and destination zones, diversionary routes during peak travel times (or from incidents) and more. He notes that as the CATT Lab evolves its tools to new datasets or requested features and functions, Plaisant’s expertise will continue to help advance the usability and usefulness of the lab’s tools. 

 –Excerpted from a UMIACS article  by Melissa J. Brachfeld and an NPR interview with  Mike Pack from the CATT lab