Pie, heatmap and others¶

Exercise pie-heatmap-others-annex.pdf · PDF

The match is over and the stats screen comes up. It does not show you one chart. It shows you four different things, because you have four different questions.

How did your damage split between weapons? A circle cut in slices. Where on the map did the fights happen? A map painted with colour, and the game even calls it a heatmap. What was your final score? One huge number. And who did what? The full scoreboard, a plain table.

So far you have four tools: bar, line, histogram and scatter. This lesson adds the four on that stats screen. One of them, the pie, is very easy to get wrong, so you will also build one that lies.

The pie chart: parts of one whole¶

A pie chart is a circle cut into slices. The circle is the whole, 100 %, and each slice is one part of it. It answers the question what share of the total is each part?

On the stats screen, the circle is all the damage you dealt and each slice is one weapon. The pie makes one promise that no bar chart makes: these slices add up to everything.

The idea is more than two centuries old. The Scottish engineer William Playfair is credited with the first pie chart, in his book The Statistical Breviary, published in 1801.

Here is a fair one from games_clean. In Unit 2 you created the column owners_tier, which sorts every game by how many people own it:

pie
    title Steam games by owners tier
    "Niche" : 17921
    "Mid" : 8141
    "Hit" : 623

You can say "two out of three games are Niche" without reading a single number. That is a pie doing its job.

Three conditions, or use bars¶

Remember the ranking from Visual perception: the eye reads position best, and a pie makes it compare angles, two ranks lower. So a pie has to earn its place. It does when all three of these are true:

Fail one condition and the answer is a bar chart.

flowchart TB
    A{"Are the slices parts of<br>one whole, adding up to 100 %?"}
    A -- "No" --> BAR["<b>Bar chart</b>"]
    A -- "Yes" --> B{"Five slices or fewer?"}
    B -- "No" --> BAR
    B -- "Yes" --> C{"Are the differences obvious?"}
    C -- "No" --> BAR
    C -- "Yes" --> PIE["<b>Pie chart</b>"]

You have already met a pie that fails. The 15-slice rainbow of games per release year in Visual perception broke the second condition, and time has an order that a circle cannot show.

A pie that lies¶

Now the dangerous one. Take the same pie and put publisher in it, keeping only the ten biggest:

pie
    title Wrong on purpose - the ten biggest publishers
    "Big Fish Games" : 284
    "SEGA" : 141
    "Strategy First" : 129
    "Ubisoft" : 123
    "Unknown" : 120
    "Square Enix" : 95
    "Sekai Project" : 94
    "Choice of Games" : 92
    "Dagestan Technology" : 89
    "1C Entertainment" : 85

What share of Steam games does Big Fish Games have? The pie says about 23 %. Nearly a quarter of the store.

Check it with a query:

SELECT ROUND(100.0 * COUNT(*) FILTER (WHERE publisher = 'Big Fish Games') / COUNT(*), 2)
FROM games_clean;

The answer is 1.06. Big Fish Games published 284 of the 26,685 games: about 1 %. The pie drew 23 %.

Where does the 23 % come from? Those ten slices hold 1,252 games between them, and 284 is 22.7 % of 1,252. The pie showed the share of the rows it was given, and it was given ten publishers out of 13,951.

This pie broke the first condition: ten publishers are not the whole. It also has ten slices, and one of them is the ghost player Unknown from Bar and line charts. The honest chart for this question is the sorted bar chart you built there.

The heatmap: a grid painted with colour¶

A heatmap is a grid. One dimension makes the columns, another makes the rows, and the colour of each cell shows a metric: the darker the cell, the higher the value.

On the stats screen, the two dimensions are the two directions of the map, and the colour is how much fighting happened in each spot. Nobody reads an exact number off it. You look for the hot zones.

That is the right way to read any heatmap. Colour shading is the last rank in the Cleveland and McGill table: the eye is bad at turning a shade into a number. But it is very good at seeing where the dark cells cluster.

A heatmap is for seeing patterns, not for reading values.

Here is the question: is there a season for releasing games on Steam? The two dimensions are the release year and the release month, and the metric is the number of games. From 2014 to 2018 that makes 5 columns × 12 rows = 60 cells, holding 24,115 games.

Look at the left grid first. What does the colour tell you? That 2018 is darker than 2014. The lightest cell is January 2014 (61 games) and the darkest is March 2018 (853 games).

But you knew that already. The line chart in Bar and line charts said that releases exploded. The growth from year to year is so strong that it drowns everything else, and the question about seasons is still unanswered.

Normalising: change what each colour is compared with¶

In the left grid, every cell is compared with all 59 others. In the right grid, each cell is compared only with the other months of its own year.

To normalise a heatmap is to change what each cell's colour is compared with: the whole grid, its own column or its own row.

Now the pattern appears. Read along the bottom row: January is the palest month in all five years. The darkest cell, the peak of each year, moves around:

So the honest answer is: there is a clear low season, January, and no single high one. The peak falls in the second half of the year in four of the five years, and in March in 2018.

The others: one number, one table¶

Not every question needs a chart with axes. Two of the most useful "charts" on a dashboard have none.

The big number¶

A big number is a chart that shows a single value, as large as possible.

It is the final score on the stats screen. How many games are in games_clean? There is nothing to compare, nothing to rank, no trend to read: the answer is 26,685, and the best visualisation of it is the number itself.

Think of Tufte's data-ink ratio. A big number has nothing left to erase: every drop of ink is data.

The table¶

A table shows rows and columns of exact values. It is the right choice when the reader needs to look things up.

It is the full scoreboard. Here are the seven games in games_clean with at least 20 million owners:

SELECT game, publisher, release_year, price, owners_min
FROM games_clean WHERE owners_min >= 20000000
ORDER BY owners_min DESC;

Rows with the same owners_min may come out in a different order on your screen.

Read the price column. Six of the seven biggest games on Steam are free. A table lets you see that, because it shows several columns of the same row side by side. A bar chart of owners_min would have shown seven bars and hidden the most interesting fact.

The stats screen: eight tools, eight questions¶

That makes eight tools. Each one answers one kind of question:

Read the table from the middle column. You do not start with "I want a pie". You start with "what share...?" and the question picks the tool.

Try it: four charts for the stats screen¶

All four charts use the dataset games_clean (see Resources for the data).

Chart 6 — Owners tier share. What share of Steam games is Niche, Mid or Hit?

To see the lying pie for yourself, change Dimensions to publisher and Row limit to 10, and update the chart. Look at the Big Fish Games slice, then run the query above in SQL Lab. Do not save this one.

Chart 7 — Releases by month and year. Is there a season for releasing games?

games_clean has a release_year column but no month. Add it as a calculated column, as you did in Synthetic variables.

EXTRACT(MONTH FROM release_date)::int

Chart 8 — Total games. New chart, chart type Big Number, Metric: COUNT(*). It shows 26.7k. In Customize, set Number format to ,d to get the exact figure, 26,685. Save it as UT3 · Total games.

Chart 9 — Biggest games table. New chart, chart type Table. Query mode: Raw records. Columns: game, publisher, release_year, price, owners_min. Filters: owners_min, Greater or equal (>=), 20000000. Ordering: owners_min [desc]. You should see seven rows, Dota 2 first. Save it as UT3 · Biggest games table.

Key takeaways¶

• A pie chart shows parts of one whole. It is acceptable only when the slices add up to 100 % of something real, there are five or fewer, and the differences are obvious. Otherwise, bars.
• A pie shows shares of the rows it was given. With a row limit or a filter it still looks complete: Big Fish Games has 1.06 % of the games, and a top 10 pie draws 22.7 %.
• A heatmap is a grid of two dimensions with a metric as colour. It is for seeing patterns, not for reading values.
• Normalising a heatmap changes what each colour is compared with. The data stay the same; the question changes.
• A big number shows one figure to remember. A table shows exact values to look up, and can reveal what a chart hides.
• Start from the question, never from the chart: eight tools, eight kinds of question.

Quick check¶

You now own eight tools and you know the question each one answers. The next step is the hard one: getting a question with no hint about which tool to pick. That is Choosing the right chart.