Kinetica: Naturalistic Multi-touch Data Visualization

I think what I like the most about this reading, although I think is not highlighted enough in the paper, is on the role of Kinetica towards attention guidance provided by adding the Physics component of the visualization. By adding Physics, many participants were observing how data points were clustering themselves into groups, by observing the trajectories of these points, participants became aware of a reason for data points to belong together and the groups themselves. Basically, this dynamic visualization helped reveal aspects of a visualization that normally may have gone unnoticed (I’m comparing against just showing a scatter plot with zero dynamics). However (according to the authors), this comes at the expense of being able to show only few(500) data points, which I think is not a problem, for example the user could always first sample the data set and then use Kinetica. The sample will have statistical properties similar to those of the entire data set with the added benefit of the physics manipulations. I think another aspect the paper overlooks is that of being able to show both the information the user cares about and information that does not(immediately) care about but could be of importance. For instance in Figure 1, the filter example shows both the data that is being filtered out and the one that does not pass. The last usually is overlooked by pretty much any filter in any software package I can think of, while I do not know immediately how to use that information I would think it is useful and it comes at almost no expense. One of the participants I believe mentioned this feature was very useful at identifying he/she had made a mistake with the filter which was not having the desired effect on the data. Last, this work would be awesome for doing fast data exploration, for instance, learning very quickly about an unfamiliar data set. 

I'd like to pickup Mary Beth's question about translating the application to a non-touch device. I came to the conclusion through reading the paper and looking at the demos that the multi-touch interface was not the essential part of the application that resulted in improved performance. Instead, it was the physics simulation and assorted tools. I'm reminded of browser-based before-and-after photo viewers that effectively utilize dragging back and forth to give users a sense of a scene from two different moments. This interaction is a very basic physics simulation and it is effective. I agree that there are some interactions—expanding a circular filter comes to mind—that would be difficult to replicate with a mouse, but I think a click (define center) and drag (define radius) would works almost as well.

I'm trying to think of other metaphors one could use as the basis for interacting with data. I remember the sedimentation metaphor (http://www.visualsedimentation.org/examples/sedivn/sedivn.html) that was moderately successful at showing real-time data. I also think trees and plant growth would be useful to show growth. Flow (like wind or a river) could be a metaphor to show how particular events change a pre-existing flow of data. These seem to be more opinionated metaphors than Kinetica's.

Replying to Question #2, I think that the notion of physics-derived operators along the lines of the framework created could be combined with spatial operators for interesting hybrid data analysis and visualization techniques. For instance, while visualizing tweets within Kinectica, you could create barriers to segregate spatial areas and use forces to cluster data within each spatial area. Within a spatial area you can apply either a similar wordcloud operator to generate word clouds based on which data points are present so as to quickly see how your clustering might make semantic sense. Alternatively you could apply different operators within each space such as a different topic model that is intended to model slightly different social phenomena. Thus you can create a spatial arrangement for political tweets that visualizes the top 3 mentioned topics in the political space, while capturing the most retweeted topics in the pop music/culture space. This form of visualization seems less useful for rapid analytics, but instead rapid dashboard prototyping/configuration.

Re: The authors incorporated physics-based affordances to their visualization by using collisions and a magnet tool. Can you think of other physics or natural interaction based affordances that can enhance data visualization?

There was a really cool paper at CHI 2015 that falls into this same category of 'natural' interactions with data, "Exploring Interactions with Physically Dynamic Bar Charts." They push the metaphor of natural interaction much further, but in doing so seem to amplify both the benefits and the limitations of Kinetica. Users can tap different areas of the 10x10 3D bar graph to annotate data, hide or highlight sections, and sort (I think?) data if they wish. I think there's an idea here that comes from Kinetica that is really interesting - that you can engage different types of learners using different perceptual and attentional cues to help make information and data more salient? This kind of digitally-augmented physical representation of data is really cool and maybe over time further developers can use the digital augmentation to solve some of the scale and processing limitations. 


Paper Link: dl.acm.org/citation.cfm?id=2702604
Video link: www.youtube.com/watch?v=TIKuE6GT_P4

I think there are really two points here; some of us have been talking about this in the context of making visualizations that increase our understanding of the data, which is obviously important, but there's also the educational value of simply getting somebody engaged in the data regardless of whether a specific visualization actually increases understanding. Like Mary Beth, I'm excited about this as a way to teach data literacy. There are obviously barriers to adoption, and cost is likely prohibitive for most of the educational environments where it would make the most difference, but I think there's a larger point to be made about playing with your data. While the ~kinetic component of this is interesting, I think the most interesting concept is simply the freedom it gives users to explore data in a semi-freeform way. I agree with Alex in that this is quite separable from the kinetic approach, but I would go further. What types of interfaces or apps could be designed that help students learn to enjoy playing with data?

This paper presents a new data visualization system that enables the user to make sense of data more easily by introducing physics-based affordances for the data, at the cost of limited scalability on data size (~500 data points). For the purpose of quick figuring out relations and trends in a dataset, this tool is quite useful. And as Julian said, one can just use Kinetica on a sample of data to avoid the limitation in scalability. But for the use of such system in education, like Mary Beth proposed, I think it would be very interesting and important to explore if the adoption of such system (and the underlying analogy between "data" and physical object) would affect students' ability in data analysis with larger dataset.

An analogy would be, does teaching addition to toddlers as "adding one apple to another apple, you get two apples" makes it harder for them to apply addition on non-apple objects? Or a more complex example, visualization helps us learn 2d and 3d geometry more easily. But does using visualization to learn 2d or 3d geometry makes it more difficult for us to learn high dimensional geometry?

(0)
Joseph wrote: "but there's also the educational value of simply getting somebody engaged in the data regardless of whether a specific visualization actually increases understanding"

This is an interesting concept - I wonder how we might characterize the educational value here. Is the idea that this would increase the chances that a student would want to play with data in the future (effectively increasing opportunities for future learning)? Or is there an assumption that, even if the student's understanding of the underlying domain is not increasing, the student is likely to come away with an increased understanding of domain-general properties of data?

(1)
Above, Anna wrote "...how much do I even really gain from understanding data better?". I share her skepticism about many existing visualization tools that are intended to augment human decision making in real world settings:
I would be very interested in seeing empirical studies that attempt to examine whether and how data visualization tools -- especially those intended for use in real-time, or over short time intervals -- actually influence human decision making in complex/noisy environments (e.g. social decision making). ...I've probably phrased this a bit too broadly. But does anyone have recommendations?

Anna also wrote:
"But as we accumulate more and more data, how relevant is data visualization .... instead of trying to make sense of big data, should we be distributing the understanding of small subsets of big data? I feel like these are to a certain extent ridiculous/impractical questions, but they're also kind of real to me."

^ I think about this all the time too... To me, one of the main purposes of relatively low-level data visualizations (e.g. visualizations presenting data that have not been *heavily* pre-processed and/or pre-interpreted -- pushing more of the interpretation work to the user) is to enable inferences that humans are likely to make (given their rich and structured prior knowledge, which contributes to their inductive reasoning advantages over machines in many learning, reasoning, and decision-making contexts), but that automated pre-processing methods are likely to miss.
Another reason (strongly linked to the last point), is to enable more flexible use of data (e.g. for different decision-making contexts) via a single visualization. Pre-processing may remove much of the work of interpretation, while also constraining the way the data can be used. And similarly, visualizing lower-level data may promote trust between the user and the system (even if the presentation of lower-level data does not ultimately result in different inferences or decisions than would have been reached via presentation of more pre-processed data).

...In cases where it may be desirable to visualize less pre-processed data, I'd agree with Julian that it would probably just make sense to sample. ...But I'm really interested in this idea of distributing understanding of small subsets of a larger dataset. Can anyone think of a contexts where this might be beneficial (above and beyond presenting one or more samples [not simultaneously] to an individual?)


(2)
A while back, I saw an interesting app in the NYT that asked users to draw their best guess of the curve (from a fairly broad space of functional forms) that best characterizes some real-world statistical relationship between two variables:
www.nytimes.com/interactive/2015/05/28/upshot/you-draw-it-how-family-income-affects-childrens-college-chances.html

I wonder whether there have been any empirical studies on the effects of this use of visualization tools as a instructional devices (i.e. having users reveal [or formulate] their preconceptions about some statistical relationship, and then receiving reasonably detailed feedback about how their preconceptions differ from the measured reality)?