THINK FIRST·CODE LATER

← Interactive Information Visualization
Chapter 5 · Week 5

Marks, Channels and Perception

Before You Start: What You Must Be Able to Do

Before the questions, make sure you can: describe any chart as marks plus channels; distinguish identity channels (what/where) from magnitude channels (how much); apply the expressiveness and effectiveness principles; rank channels for quantitative, ordinal and categorical data and justify a design with them; use Stevens' power law to predict how area, brightness and length are perceived; explain separability, pop-out and grouping; explain why bars need a zero baseline and compute a lie factor; and critique an encoding (yours or an AI's) in these terms.

The Big Idea

Every chart, from a bar chart to a network diagram, is built from two things: marks (geometric primitives — points, lines, areas) and channels (visual properties of marks — position, length, colour, size, shape) that encode data attributes. Choosing which attribute goes to which channel is the core design decision, and it is not a matter of taste: decades of perception experiments show that people read some channels far more accurately than others. An AI may choose a pie chart because it is common; you will choose the encoding because the evidence says it works for the task.

Marks and channels

  • Marks by dimension: points (0D), lines (1D), areas (2D). In networks, marks also express relations by connection (lines between nodes) or containment (nested areas, as in treemaps).
  • Channels control a mark's appearance: horizontal and vertical position, length, angle/tilt, area/size, colour hue, luminance, saturation, shape, curvature, motion, texture.

A scatter plot: point marks, x position = temperature, y position = PM2.5, hue = district. A bar chart: line/area marks whose length encodes a quantity and whose position separates categories.

Two kinds of channels

  • Magnitude channels (ordered): how much? — position, length, angle, area, luminance, saturation. For quantitative and ordinal attributes.
  • Identity channels (categorical): what? where? — spatial region, colour hue, shape, motion. For categorical attributes.

Expressiveness and effectiveness

From Mackinlay's work (1986), restated by Munzner:

  • Expressiveness: the visual encoding should express all of, and only, the information in the attribute. Ordered data needs an ordered channel; categorical data must not be shown with a channel that implies an order. Encoding district names with size says "Ouhai is bigger than Lucheng" — false. Encoding PM2.5 with shape cannot express magnitude at all.
  • Effectiveness: the most important attributes get the most effective channels.

Channel rankings

Munzner's ranking, from most to least effective:

Magnitude channels (ordered attributes) Identity channels (categorical attributes)
1. Position on a common scale 1. Spatial region
2. Position on unaligned scales 2. Colour hue
3. Length (1D size) 3. Motion
4. Tilt / angle 4. Shape
5. Area (2D size)
6. Depth (3D position)
7. Colour luminance, saturation
8. Curvature, volume

This is why the bar chart and the dot plot (position/length on a common scale) beat the pie chart (angle/area) for comparing values, and why a heat map (luminance) is good for patterns but poor for reading exact values.

The evidence: graphical perception

Cleveland and McGill (1984) asked people to judge the ratio between two marked values in different chart types. Errors were lowest for position along a common scale, then position along non-aligned scales, then length, direction and angle, then area, then volume and colour. Heer and Bostock (2010) replicated the experiments with crowdsourced participants on Amazon Mechanical Turk and found the same order — and showed that crowdsourcing is a valid way to run perception studies (Chapter 14).

Stevens' power law describes how perceived intensity grows with physical intensity: perceived = k · stimulusⁿ. Exponents measured by Stevens include length 1.0, area 0.7, depth 0.67, brightness 0.5, saturation 1.7 and electric shock 3.5.

  • Length (n = 1): a bar twice as long looks twice as long — accurate.
  • Area (n = 0.7): a circle with twice the area looks only 2^0.7 ≈ 1.62 times larger — underestimated. Bubble charts make large values look smaller than they are.
  • Brightness (n = 0.5): four times the luminance looks only twice as bright.
  • Saturation (n = 1.7): differences look larger than they are.

Accuracy, discriminability, separability, pop-out, grouping

  • Accuracy — how precisely a value can be read (the ranking above).
  • Discriminability — how many distinct values a channel can show: people reliably tell apart about 6–9 hues and around half a dozen shapes. With 12 districts, hue fails; use position (small multiples) instead.
  • Separability — some channels can be read independently (separable: x position and hue), others merge into one percept (integral: width and height are seen as area; two colour axes are seen as one colour). Size and hue interfere somewhat (small marks show hue poorly).
  • Pop-out — some channels are processed pre-attentively: one red dot among blue ones, or one tilted line, is found instantly regardless of how many distractors there are. Combinations (red and square among red circles and blue squares) usually do not pop out.
  • Grouping (Gestalt principles) — proximity, similarity, enclosure, connection, continuity: viewers see items as belonging together when they are close, look alike, are enclosed or connected.

Weber's law: the just-noticeable difference is proportional to the magnitude of the stimulus. Comparing the lengths of two unaligned bars is hard because the eye must compare total lengths; aligned bars on a common baseline turn the comparison into a judgment of the small difference at the top — one reason common scales win.

Bars need zero: the lie factor

A bar encodes its value by length, so the axis must start at zero. Edward Tufte's lie factor = (size of the effect shown in the graphic) / (size of the effect in the data). With the axis starting at 40, PM2.5 values 42 and 44 (data: +5 %) are drawn as bars of length 2 and 4 (+100 %): lie factor 21. Lines and points encode by position, not length, so a line chart may use a non-zero baseline to show variation — as long as the axis is clearly labelled.

One dataset, many encodings

The same CityPulse table (date, district, PM2.5) can be a line chart (trend over time per district), a bar chart of means (compare districts), a strip or box plot (distribution), a heat map with districts × days (patterns, luminance) or small multiples (one panel per district). Each makes a different task easy — that is the point of the encoding studio in this chapter's labs: choose by task (Chapter 3) and by channel ranking, not by habit.

Industry spotlight

Design systems at large companies (e.g., the data-visualization guidelines of IBM Carbon, Google Material or the UK Office for National Statistics) codify these rules: bars start at zero, no 3D, no pie charts with more than a few slices, colour only for categories that matter. They exist because perception research translates directly into fewer misread reports.

Research spotlight

Do AI models "see" charts like people? Studies of multimodal language models answering questions about charts find that they read values from position reasonably but struggle with area, angle and colour judgments and with cluttered designs — a machine echo of Cleveland and McGill. Designs that are effective for humans also tend to be more readable for AI chart-question-answering systems (Chapter 12).

Key takeaways

  • Charts = marks (points, lines, areas; connection, containment) + channels.
  • Magnitude channels for ordered data; identity channels for categories. Expressiveness: all and only the attribute's information. Effectiveness: best channels for the most important attributes.
  • Ranking: position (common scale) > position (unaligned) > length > angle > area > depth > luminance/saturation > curvature/volume; identity: spatial region > hue > motion > shape.
  • Stevens: length n = 1 (accurate), area 0.7 and brightness 0.5 (underestimated), saturation 1.7 (overestimated).
  • Discriminability (≈ 6–9 hues), separability (integral vs. separable), pop-out, Gestalt grouping, Weber's law.
  • Bars start at zero; lie factor = effect in graphic / effect in data.

Ready? Close the notes and practise.

30 questions. Predict the output before you check — that is the skill the exam measures.