Visualization is not decoration. It is cognitive engineering. The chart is an interface between the data and the mind. The interface determines what the mind perceives. A bad interface hides the signal. A good one makes the signal impossible to ignore.
Visualization is the presentation of data in a form that humans can perceive patterns, anomalies, and relationships. A table of 50,000 rows is data. A line chart of the same data is information. The difference is the visual encoding — position, length, color, shape, orientation — that maps data attributes to perceptual channels. The mapping is a design decision. The design determines whether the viewer sees the signal or the noise.
The theory of visualization is the theory of human perception. Jacques Bertin's Semiology of Graphics (1967) identified the visual variables: position, size, value (lightness), texture, color, orientation, shape. Each variable has different perceptual properties. Position is the most accurate — humans can compare positions along a common scale with high precision. Color is less accurate — humans perceive color categorically, not continuously. The design principle: map the most important data attribute to the most accurate perceptual channel. Position for quantities. Color for categories. Shape for nothing important — it is the least accurate channel. The principle is violated in most dashboards, which use color for quantities and position for nothing.
Edward Tufte's The Visual Display of Quantitative Information (1983) defined the principles of graphical excellence: show the data, induce the viewer to think about the substance, avoid distorting what the data has to say, present many numbers in a small space, make large datasets coherent, encourage the eye to compare different pieces of data, reveal the data at several levels of detail, serve a clear purpose. Each principle is a constraint. Each constraint improves the result. Tufte's most famous rule: maximize the data-ink ratio — the proportion of ink used to present data compared to total ink used in the graphic. Erase non-data-ink. Erase redundant data-ink. The erasure is the discipline.
A hospital infection control dashboard illustrates the stakes. Raw data: 50,000 surgeries, 200 columns. The dashboard visualizes infection rates as a control chart — a line graph with upper and lower control limits, each theater as a separate line. When Theater 7 exceeds the upper control limit, the point turns red. The red point triggers a signal. The signal triggers an investigation. The investigation finds a changed sterilization protocol. The protocol is reverted. Infections decline. The visualization saved lives. The visualization worked because the design made the anomaly perceptible. A table of 50,000 rows would not have. The table is data. The chart is a decision support system.
See: Edward Tufte, "The Visual Display of Quantitative Information" (Graphics Press, 1983). Jacques Bertin, "Semiology of Graphics" (1967, English translation 1983). Leland Wilkinson, "The Grammar of Graphics" (Springer, 1999) — the theoretical foundation of ggplot2 and Vega-Lite.
This post is part of a series on The Many Vs of Data, originating from Doug Laney's 2001 Gartner note. Each V names a dimension of why data is hard.