Gordon will be presenting the paper tomorrow at the Design Process and Persuasion session, at 2:20PM.
We’ll be doing a series of posts about this work that will follow the structure of the talk. This post is an adapted version of the first act of the talk.
Let’s get started
This is a theory paper about a mathematical approach to understanding, critiquing, and improving visualizations.
This isn’t a typical vis theory paper, so we didn’t think it would be well served by a typical talk. Rather than structure the talk directly around the paper,
I’ll give an alternate path into the ideas of the paper, by focusing on some example visualizations, and the questions we can ask about them.
I’ll also talk at a high-level about the math in our “Algebraic design”, there are more details in the paper. Some related work will be highlighted as it comes up.
The basic idea
We want to rethink the structure of theories of visualization. Instead of saying “Dataset is X, so then vis should be Y”, and talking about taxonomies of data, tasks, and methods, maybe we can say “we can X the data; can we Y the picture?”
So it’s less about nouns, and categorization, and more about verbs, and transformations.
Our approach is actually entirely about transformations, differences, and changes both in the data, the interesting or important ways the data to be visualized could have been different, and in the visualization, changes which are clearly visible when they connect to perceptual channels and to affordances.
A designer using our approach will constantly be asking:
The basic question
“Are the important data changes well-matched with obvious visual changes?”
We think this one question is an incisive tool for exposing how a given visualization method does or does not work for a given task, and it helps to explain why.
We developed three algebraic visualization design principles, by considering different facets of this one question.
But let’s start with an example.
2014 Election Results in Virginia
Let’s say we have a visualization of election results. Here we are showing the results of 2014’s senate race in Virginia, as reported by the New York Times.
We are all used to saying that a visualization of election results that uses only full-red and full-blue is not a very good visualization.
Spefically, we tend to say that is not as good as a visualization which uses a graded red-purple-blue colormap.
In turn, we sometimes hear “what you should really do is use a double-ended colorscale instead”, which transitions from red to blue through gray, instead of directly purple.
But we want to go a little deeper: we would like to be able to explain in what sense these visualizations are better than one another. The way we are going to tease apart these explanations is by thinking about what would happen to the image if we were to change the data in some way. If we transform the data, what’s the corresponding change in the picture? Let’s take it from the beginning.
Consider the first colormap again.
We all have some intuition for why this is not a great visualization, right? What is going on is that this picture shows us nothing about how close each precinct’s results are. In the way of thinking we will discuss in this talk, we need to find a change in the dataset that works as a “witness” for this problem.
Let us look at a scatterplot of the precinct result as a proxy for the dataset, and let us imagine that this was a much tighter election.
Now, what we will do is that for every precinct, we will push the results 80% of the way to a tie, and look at happens to the image.
Now, we know we happens here: the visualization is exactly the same! If democrats or republicans won by 10%, 1% or 0.1%, that makes no difference for the colormap. This is not great. In our paper, we call any change on the data that does not change the picture a confuser for the visualization. This issue of visualizations failing to be injective has been pointed out in the literature by Ziemkiewicz and Kosara. Here, though, we can go a bit further, and characterize the nature of the injectivity failure as changes in the data (in this case, making the election closer).
But now that we have described the problem with the vis entirely in terms of the data, we can take this description of the problem and see what it does to other visualizations.
Specifically, we can just plug this exact change of data (making the elections closer) into other visualizations, like the red-purple-blue colormapping, and see what happens with the image.
In the case of the red-purple-blue visualization, we see that all precincts in the picture move towards a central color. So the confuser for the red-blue map is not a confuser for the red-purple-blue map: that is one specific sense in which the blue-purple-red map is better than the blue-red map. New terminology: In this talk we will call the changes in the data “alpha”s, and “omega”s will be the corresponding changes in the images.
We just showed one alpha that lets us compare the visualizations in some specific way.
But we are not limited to a single alpha, and as we will see, picking different alphas is a great way to explore different aspects of a visualization. So let us now transform the data differently, by flipping the outcomes. Every vote for the republicans now count for the democrats and vice-versa.
The red-blue visualization does exactly what you expect: blue states become red states, and vice-versa. That’s good.
But now, let’s look at what happens for this same transformation, except using the red-purple-blue visualization.
For the precincts with extreme outcomes, it’s easy to see what happens: like the red-blue colormap, reds become blues and vice-versa. But let’s pay attention to the tight results. In this case, we have a slightly-reddish-purple turning into a slight-blueish-purple, and vice-versa. That’s not so great: whether or not a color will become more reddish or less reddish under this omega depends on being to either side of a particular shade of purple, a distinction which our eyes are not very good at making.
More specifically, it’s hard to see, without flipping the data back and worth, which shade of purple will stay the same. These are important because when you think about the data values, those that do not change under the “inversion” alpha are precisely the tied precincts. So what we are effectively doing is using this alpha to model ties, and noticing that the omega is not great at picking out these ties.
Notice that talking about omegas is necessarily talking about the human vision system. So discussions of omega need to happen based on our scientific knowledge and models of the human vision, which are central issues in perceptual research in vis.
In the case of this red-purple-blue example, we say there was a failure in that the visualization caused a jumbler for the notion of a tie in a precinct. This is a different kind of problem than a confuser: the data change is reflected in the picture, but it is done in a way that our visual system is bad at perceiving.
In contrast, consider a red-gray-blue election map.
In this visualization, the fixed point for the omega corresponding to the inversion alpha is gray.
The inversion keeps the saturation of the color the same, and a tied race corresponds to a completely desaturated color. In this visualization, then, the inversion alpha gets mapped to a categorical change in the hue, which happens to be a no-op when there is no saturation. When these hues are opposing, this omega uses a structure that our eyes are good at capturing. We can also examine the red-gray-blue election map with the first alpha, that of making elections closer.
Analogously, the tranformation that makes the election closer gets mapped to another structure our eyes are relatively good at capturing: reduction in saturation. At the same time, the hues do not change: blues stay blue, and reds stay red. (Note, that in practice one has to worry about simultaneous contrast, and this present analysis we are showing you does not account for that. Nothing stops us from performing an algebraic analysis of simultaneous contrast issues, though!)
So now you’ve seen the basics of our algebraic process. We identified some interesting possible changes in the data, and then studied how the visualization method maps those to changes in the final image and based on that, we could concretely criticize which color scales work better for what thing.
For a theory of visualization design, though, we want to formalize and generalize the design strategy we just used. This is what the next post will be about.