Simple Changes in Chart Arrangements Make Some Messages More (or Less) Obvious
A brief summary of our newly published research paper on visualization affordances.
I’ve been wanting to share the ideas behind this research for a long time, and we finally have a paper out that documents the whole thing. Well ... not the whole thing because the amount of thinking that went on this is enormous, and what’s in the paper is the tip of the iceberg. Here, I’ll briefly summarize what we describe in the paper.
Before I move on, I want to highlight that the main person behind this research is my Ph.D. student Racquel, who has brilliantly developed many of these ideas and studies and who “sparred” intellectually with me (and my colleague Steven) for at least a couple of years in order to develop the ideas contained in the paper. (Side note: If you happen to be at the IEEE VIS conference as you read this make sure to attend her talk!)
The main idea explored in the paper is the fact that the message a chart tends to communicate depends, in part, on graphical properties that “suggest” what information the chart contains. In other words, charts use graphical properties that influence the way charts are read and interpreted and, as a consequence, influence what message is extracted from them.
To capture this idea, we borrowed the concept of affordance from user interface design. Affordance is the idea that how an artifact (digital or physical) appears suggests how it can be used. Drawing a parallel, we use the concept of visualization affordance to communicate the idea that some graphical properties suggest how a chart can be read and, as a consequence, what type of message it is designed to convey.
(Note: the term affordance, introduced by Donald Normal in the world of UI design, is not entirely correct. We decided to use it anyway because it is commonly found both in UI design references and in a good number of publications in data visualization.)
Another way to look at this problem is that if you want to communicate a given message, some graphical properties will make a chart more or less obvious according to what extent the properties match the message you want to communicate.
To make progress in this area, two elements are necessary:
Identify which properties may have such an effect
Verify experimentally whether they have the expected effect
Our paper moves the first steps in this direction by using bar charts, arguably the simplest and most common chart that exists. We identified a number of arrangements we hypothesize have an effect on message obviousness and then verified our intuition with a series of experimental studies.
Bar arrangements
For this study, we identified three main graphical properties:
Spacing. Space between groups of bars.
Coloring. Bars colored with different colors.
Ordering. Bars ordered according to their length.
Partitioning. Bars partitioned into segments (i.e., stacked bars).
The main idea is that these four properties have an effect on whether a given bar chart is a good match for a given type of message. But, if we do not define what we mean by “type of message,” we can’t really study this effect. For each of these arrangements, we have corresponding effects on messages; that is, when we use different spacing, coloring, sorting, or partitioning strategies, there is an effect on what kind of message the chart conveys.
Type of messages
We have identified three main types of message effects that are affected by the properties mentioned above:
Grouping. It communicates the idea that certain elements form groups and as such encourages the elements within a group to be part of a category or entity. We also expect the grouping to suggest comparisons between the groups. In other words, if a visualization contains groups, the visualization will promote a reading where comparisons between groups come naturally.
Ranking. It communicates that the elements are ordered, and some are the top or bottom ones according to a given metric or criterion. If a visualization has sorted elements it will promote a focus on the ranking of the elements (first, second, third, etc.), including qualitative interpretations such as “best,” “worst,” etc.
Proportions. It communicates the idea that individual elements form a group, including interpretations in terms of percentages of a total, such as, “A is about 30%, B is about 20%, and the rest is about 50% of the total.”
The experiments
We designed the experiment using the same strategy for each type of arrangement and message.
We present four charts for each of the arrangements we have outlined above. Some include the affordance effect we want to study, and some do not. For example, when we study the effect of ordering, we have bar charts that are sorted and bar charts that are not, like in the figure below.
When we study partitioning, we have plots with stacked bars (that is, with partitions) and plots with grouped bars, like in the figure below.
For each arrangement type, we designed a set of messages that express or do not express the messages we believe are a good match for the affordances the plots express. For example, when we study the expression of the ranking concept, we have messages such as, “Company J sold the largest number of items,” which focuses on ordering, and “In total, Companies AEF sold a smaller number of items than companies JKM,” which does not.
For each message, we ask which of the four charts makes the message “more obvious” to our participants, and then we verify whether there is a good match between the affordance the charts have been designed to express and its matching concept; that is, we count how many times the participants use the chart that expresses the affordance we hypothesize is a good match.
The results show a strong effect of affordance-message pairing; that is, people tend to choose the chart that has a stronger correspondence between the message and the affordance. I’ll not include all the results here. It’s rather dense, but you can look at the paper to see that in most cases, the participants responded as expected.
So what?
You may ask yourself, “Interesting exercise, and now what? Why should I care? What you found is pretty obvious, no?”
You should care because this is a first step towards something that we do not really have in visualization, which is a catalog of graphical properties common in charts that can impact what type of message a chart conveys.
While visualization theory often focuses on the idea of efficiency and effectiveness in extracting “quantities” from a visual representation, here we focus on what kind of properties have an impact on how charts are interpreted by the readers. While we all intuitively use some of these properties to convey a message, often, charts are hard to interpret because there is a poor matching between what the designer wants to express and what the charts convey. Studies like this one can help guide and rectify these problems.
We hope that if we have a better sense of what these properties are and what kind of basic messages charts convey, we can have a more explicit organization of this knowledge for visualization design. In turn, this can be a tool for designing and evaluating visualizations as well as a pedagogical framework to teach learners how to create more effective visualizations.