Encoding data into visual attributes is one of the fundamental processes of data visualization. The challenge is to find the right encoding for the data to be represented. In fact, errors in the choice of encoding is the most common error in creating a visualization.
Several researchers have created rankings indicating which visual attributes are best for which types of data. But they don’t all list the same visual attributes. They don’t all have the same rankings. They don’t use the same scoring model. At the same time, it would be very interesting to bring all the ranking models together to compare them and create some kind of meta rank.
Meta-table of Visual Attribute Rankings
The table immediately below brings together the rankings of several researchers:
Ber67 – Jacques Bertin, Semiology of Graphics
Mac86 – Jock MacKinlay, Automating the Design of Graphical Presentations
Mac06 – Jock MacKinlay, Designing Great Visualizations
Mac96 – Alan MacEachren, How Maps Work
Maz09 – Riccardo Mazza, Introduction to Information Visualization
Under each author is the original score/rank. Bertin simply provides yes/no. Jock provides a rank diagram, which is converted into a rank 1-n. MacEachren and Mazza provide ordered rankings such as poor-marginal-good or no-limited-suitable. These are converted into uniform 1-9 scores (e.g. Bertin yes = 1, no = 9). The average score is the sum of the individual scores divided by the number of responses; lower score is a good encoding, higher score is a poor encoding. Attributes scored by only a single researcher are not included in the final list of average scores.
- All authors are fairly consistent in rankings across quantitative variables. Use position first, use size second. Nothing else ranks even close to these two attributes – angle and brightness are in a distant third and fourth place. Don’t use shape or grouping/containment.
- For ordered data, things shift. Position still ranks top but brightness is second. Saturation and texture come in next. Texture? I can’t think of many visualizations (any?) that use texture for ordered data, except possibly some examples in Bertin’s book.
- For categorical data, the visual attributes are different: texture, color hue and position all have nearly the same top rank score. Shape is only marginally below the top 3, mostly because Jock changed it’s rank from 7 in 1986 to 2 in 2006.
Jock didn’t include some variables in 2006 and the remaining ranks were stretched out to the 1-9 range bring his 2006 rank of 3 for color down thereby pulling color down. Don’t use size, brightness nor saturation for category data – the ordering of these attributes presumably is perceptually hard-coded and we can’t not see them as ordered.
- Across the board, position always ranks top – think how small multiples nest extra positional encodings inside a larger positional encoding.
The model has a few flaws:
- Jock MacKinlay gets two votes since two of his attribute lists are referenced (and why did Jock revise his rankings in 2006?)
- Some other attribute ranking lists should be added here, such as Illinsky’s list and Munzner’s list.
- How many rankings should be required in order to have a valid average score?
- Positional separation always ranks highly.
- For quantitative data, use position and size
- For ordered data, use position and brightness (and maybe texture saturation)
- For categoric data, use position hue, texture and shape.