Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods
William S. Cleveland; Robert McGill
Journal of the American Statistical Association
This paper is a little different from the previous papers I’ve read. It does not outline new optimization methods or review machine learning algorithms. This paper just looks out how we display data, which has become a huge field in the most recent years. Cleveland and McGill wrote this paper before the wave of data analysts and data scientists, back in 1984. However, there findings are still very relevant and super useful.
The paper outlines a study on perception. The authors sought to determine which charts and graphs are the most easily interpret-able and most accurate. They broke down the interpretation and structure of graphs into 10 “elementary perceptual tasks” which describe features that graphs use to separate data. The ten tasks are as follows ranked in best to worst:
1.position along a common scale
2. position along non aligned scales
3. Length, Direction, Angle
4. Area
5. volume, curvature
6. shading, color saturation
This is quite an interesting list especially when looked at in terms of todays graphical documents. One of my favorite visualizations is the cholorpleth maps which rely almost entirely on color saturation, however this task fared the worst! Keeping these tasks in mind the authors iterate through common graphs. Some scored better than others, Bar charts topped most of their tests, while the widely hated pie chart scored toward the bottom.
This paper is full of graphs and charts to show their findings and examples of how some graphs fail and other succeed. A particularly interesting example is the distance between two curves. On the left the show a matrix of two curves and asked their participants to estimate the distance between the two curves at various points. On the right is the actual difference of the two curves. I found that even after reading the paper I would stumble on my perception of the two curves.
This paper is excellent. I highly suggest that anyone who makes graphs gives it a quick read. The graphs look a little dated, but nevertheless contain tons of information. It even has some recommendations on common graphs to replace with graphs that better display information.
One thing the paper does not capture is the recent trend to make graphs as pretty as possible. There is an obvious trade off that the creator must decode. Do I want to make a pretty graph that entices clicks or a utilitarian graph which conveys the most amount of information? Reading this paper brings us a little closer to a happy compromise.
One thing the paper does not capture is the recent trend to make graphs as pretty as possible. There is an obvious trade off that the creator must decode. Do I want to make a pretty graph that entices clicks or a utilitarian graph which conveys the most amount of information? Reading this paper brings us a little closer to a happy compromise.