Visual perception¶

Exercise visual-perception-annex.pdf · PDF

You are in the middle of a boss fight. You do not read your health. You glance at a bar in the corner, see that it is short and red, and drink a potion. Half a second, no reading, no maths.

That corner of the screen is the HUD (heads-up display): the health bar, the minimap, the ammo counter. It works because somebody designed it for an eye that has no time. A chart has exactly the same problem. Your reader gives it a glance, and in that glance the chart either answers the question or loses them.

In Units 1 and 2 you obtained the data and cleaned it. You now own games_clean: 26,685 Steam games, tidy and trustworthy. This unit is the last stage of the data lifecycle: represent. Before drawing a single chart, you need to know how the eye reads one.

Count the sevens¶

Try this before reading on. How many sevens are in the first grid? Count them, and notice how long it takes.

Now look at the second grid. The answer is 6, and you had it before you decided to count.

The data did not change. The digits are the same, in the same places. One visual property changed, and your eyes did the work on their own.

A preattentive attribute is a visual property that the eye detects automatically, in a single glance, before you pay conscious attention.

How fast is "a single glance"? Vision researchers use a threshold of about 200 to 250 milliseconds, a quarter of a second. That is less time than your eye needs to start moving to another spot, so whatever you detect in that window you detected without searching.

The effect has a name: pop-out. The sevens popped out of the grid.

There are many preattentive attributes. These are the four you will use all unit, each one already at work in your HUD:

Position beats angle¶

A chart turns numbers into something you can see. That translation is called an encoding: this number becomes a length, that one becomes an angle, another becomes a shade of colour.

Not all encodings are equal. Here are five Pokémon types from the class table pokemon, drawn twice. Which type has the most Pokémon? And which has the fewest?

In the pie you probably found bug, and then got stuck. Is fire smaller than electric? The slices are 59 and 61, and your eye cannot tell. In the bars you see it at once, even without the numbers on top.

This is not an opinion. In 1984 two statisticians, William Cleveland and Robert McGill, published the paper Graphical Perception in the Journal of the American Statistical Association. They showed people simple graphics, asked them to judge the values, and measured the errors. Then they ranked the encodings, from the one people read most accurately to the one they read worst:

The higher in the table, the more precise the reading. Bars sit at rank 1: all of them start on the same line, so you only compare where they end. A pie asks you to compare angles, rank 3.

So the bar chart is your default answer. A pie has to earn its place, and in Pie, heatmap and others you will see exactly when it does.

Colour means something¶

Colour is the strongest pop-out tool you have. That is exactly why you must not waste it.

Colour has three honest jobs:

Decoration is not a job. If every bar has a different bright colour "to make it lively", the colours say nothing, and the reader still tries to work out what they mean.

Visual hierarchy¶

Visual hierarchy is the order in which the eye moves through a picture: what it sees first, second and third.

In a chart the eye follows a predictable path: the title, then the biggest or most coloured mark, then the rest. You control that path.

Decide what must be seen first, and make it the biggest thing or the only coloured thing. Everything else goes quiet: grey, thin, small.

Sorting is hierarchy for free. Look at the same five bars, first in the order the database returns them and then sorted:

%%{init: {"themeVariables": {"xyChart": {"plotColorPalette": "#4051b5"}}}}%%
xychart-beta
    title "Steam games by price band"
    x-axis ["15 to 30", "30 or more", "5 to 15", "Free", "Under 5"]
    y-axis "Games" 0 --> 12000
    bar [2367, 565, 9046, 3094, 11613]

%%{init: {"themeVariables": {"xyChart": {"plotColorPalette": "#4051b5"}}}}%%
xychart-beta
    title "Steam games by price band"
    x-axis ["Under 5", "5 to 15", "Free", "15 to 30", "30 or more"]
    y-axis "Games" 0 --> 12000
    bar [11613, 9046, 3094, 2367, 565]

Same data, same chart type. In the sorted version the answer to "which band is the biggest?" is the first thing on the left, and the ranking is readable without comparing bars one by one.

Data-ink: erase what shows no data¶

In 1983 the American statistician Edward Tufte published The Visual Display of Quantitative Information. Its most quoted rule is five words long: "Above all else show the data."

The data-ink ratio is the share of a chart's ink that shows data. The rest is ink that could be erased without losing a single number.

Tufte gave the useless ink a name: chartjunk. The usual suspects:

Critique a chart in four questions¶

Here is a chart built from games_clean: the number of games released each year, as a pie. Look at it for a moment before reading on.

pie
    title Games per release year
    "2004" : 53
    "2005" : 58
    "2006" : 112
    "2007" : 143
    "2008" : 183
    "2009" : 334
    "2010" : 340
    "2011" : 364
    "2012" : 426
    "2013" : 556
    "2014" : 1352
    "2015" : 2544
    "2016" : 4503
    "2017" : 6871
    "2018" : 8845

Every chart you make or meet from now on gets the same four questions:

Now the same 15 numbers with the rules applied: position instead of angle, years from left to right, one colour.

%%{init: {"themeVariables": {"xyChart": {"plotColorPalette": "#4051b5"}}}}%%
xychart-beta
    title "Games released on Steam per year"
    x-axis ["2004", "2005", "2006", "2007", "2008", "2009", "2010", "2011", "2012", "2013", "2014", "2015", "2016", "2017", "2018"]
    y-axis "Games" 0 --> 9000
    bar [53, 58, 112, 143, 183, 334, 340, 364, 426, 556, 1352, 2544, 4503, 6871, 8845]

What can you read now that the pie hid? The explosion after 2013: 556 games in 2013, 8,845 in 2018. The pie had that story inside it and no way of telling it.

Try it: your first chart¶

Time to build one. The question: how many Steam games are there in each price band? You need the table games_clean from Unit 2 (see Resources if you need the data again).

Start every chart name of this unit with UT3 ·. In Unit 4 you will search for that prefix to build a dashboard.

The chart must show what this query returns in SQL Lab:

SELECT price_band, COUNT(*) AS n,
       ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER (), 1) AS pct
FROM games_clean GROUP BY price_band ORDER BY n DESC;

When you hover over a bar, the box rounds the number (11.6k). The exact figure is in Results, under the chart.

Key takeaways¶

• Preattentive attributes (colour, size, position, shape) are detected in a glance, in under about a quarter of a second. Pop-out works only when the highlighted thing is the exception.
• Cleveland and McGill (1984) ranked the encodings: position on a common scale is read most accurately; angle, area, volume and colour shading come after it. Bars are the default; a pie must earn its place.
• Colour has three honest jobs: highlight, category, state. Never decoration, and never red versus green alone.
• Visual hierarchy: decide what must be seen first and make it the biggest or the only coloured thing. Sorting bars by value is hierarchy for free.
• Data-ink ratio (Tufte, 1983): erase the chartjunk, the ink that shows no data. Titles and axis labels are not junk.
• A good chart is one where the reader finds the answer in seconds. Use the four critique questions on every chart, starting with your own.

Quick check¶

Now you know what the eye catches in half a second. Next, the two charts you will use most, and the one question that decides between them: Bar and line charts.