22 Data Visualisation: The Absolute Minimum You Must Know
A chart is not decoration — it’s an argument, and it can lie as fluently as a sentence can. Effective visualisation rests on two skills, both on this page: matching the chart to the question, and refusing the standard ways charts mislead. (For the mechanics of drawing, see TAMYMN-Matplotlib.md; this page is about what to draw.)
22.1 Start With the Question, Not the Data
Every chart answers exactly one of four questions. Name your question first and the chart type falls out — this table is nine-tenths of visualisation:
| Your question | The answer is a… | matplotlib verb |
|---|---|---|
| Comparison — which is bigger? | bar chart | ax.bar |
| Distribution — how are values spread? | histogram | ax.hist |
| Relationship — do these move together? | scatter plot | ax.scatter |
| Trend — how does it change over time? | line chart | ax.plot |
If you can’t say which question you’re asking, no chart type will save you — and if you ask an AI for “a visualisation” without naming the question, it can’t know either.
22.2 Why Pie Charts Mislead
A pie chart encodes values as angles and wedge areas — and human eyes are bad at comparing angles. Is the 23% slice bigger than the 26% one? You genuinely cannot tell; turn the same numbers into bars and the answer is instant, because bars encode value as length, which eyes compare superbly. A pie is defensible only for “one obvious slice versus the rest of the whole” — beyond two or three wedges, or for any comparison that matters, use a bar chart. (A pie of values that don’t even sum to a meaningful whole — say, average temperatures by city — is not a chart, it’s an accident.)
22.3 Label Your Axes — With Units
An unlabelled axis makes the chart unfalsifiable: “sales” of what, in which units, over what period? The non-negotiable minimum is an x-label, a y-label with units, and a title — and the strongest titles state the takeaway (“Rent rose 12% in 2025”), not the topic (“Rent data”). In matplotlib that’s three set_* calls, and because Axes are objects you can verify them in tests:
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots()
>>> _ = ax.bar(['Perth', 'Sydney'], [402, 415])
>>> _ = ax.set_ylabel('Median weekly rent (AUD)')
>>> _ = ax.set_title('Sydney rents edge out Perth')
>>> ax.get_ylabel()
'Median weekly rent (AUD)'
>>> float(ax.get_ylim()[0]) # bars start from zero — an honest comparison
0.0
>>> plt.close(fig)22.4 The Lie Factor: Truncated Axes
Edward Tufte’s lie factor is the size of the effect shown divided by the size of the effect in the data; an honest chart scores about 1. The classic inflator is the truncated axis: rechart Perth (402) vs Sydney (415) with the y-axis starting at 400 and Sydney’s bar towers at four times Perth’s height — a visual 300% shouting about a real 3%. The rules that keep you honest:
- Bar charts start at zero, always. A bar’s length is its value; cutting the axis cuts the truth. (Matplotlib does the right thing by default, as the doctest above shows — truncation requires a deliberate
ax.set_ylim(400, ...).) - Line charts may zoom — a trend lives in the slope, not the bar length — but say so visibly when the baseline isn’t zero.
- Read others’ charts with the same eye: axis starts, doubled y-axes, and 3-D perspective are where the lies live.
22.5 Directing the Machine
An AI will happily generate a chart for any data you paste; only a prompt that names the question type, the encoding, and the honesty rules gets you the right chart. You are the one who knows what question the audience is asking.
Vague:
"visualise this sales data"Informed:
"This is a trend question: monthly revenue over 24 months. Line chart (not bars),
x label 'Month', y label 'Revenue (AUD, thousands)', title stating the takeaway.
The y-axis needn't start at zero, but annotate the baseline if it doesn't.
Use the fig, ax = plt.subplots() idiom from TAMYMN-Matplotlib.md."22.6 Spot the Confabulation
An AI assistant polishes a chart comparing two branches’ quarterly sales (Branch A: $1.02M, Branch B: $1.07M):
ax.bar(['Branch A', 'Branch B'], [1.02, 1.07])
ax.set_ylim(1.00, 1.08) # zoom in so the difference is easier to see —
# without this the bars look almost identicalWhat’s wrong?
“The bars look almost identical” is the truth — the branches differ by about 5%, and an honest bar chart is supposed to look that way. Truncating the axis to (1.00, 1.08) makes B’s bar roughly three times A’s height: a lie factor near 70. The comment even confesses the motive. If the 5% gap genuinely matters, keep the zero baseline and annotate the bars with their values, or plot the difference explicitly — don’t redraw the ruler until the answer looks impressive.
22.7 Where to Practice
- From Data to Viz — a decision tree that walks from “what kind of data do I have?” to a recommended chart, with the classic caveats (including the pie-chart problem) explained on the way. Free, no signup.
- WTF Visualizations — a stream of real published charts that mislead. Practise naming why each one fails; it trains the same eye you’ll use on your own drafts.
22.8 Quick Reference
| Rule | In short |
|---|---|
| Name the question first | comparison / distribution / relationship / trend |
Comparison → ax.bar |
lengths, from zero |
Distribution → ax.hist |
shape, spread, outliers |
Relationship → ax.scatter |
one point per observation |
Trend → ax.plot |
slope carries the message |
| Pie charts | angles are unreadable — use bars beyond 2–3 wedges |
| Labels | x, y with units, takeaway title — or it’s a rumour |
| Lie factor ≈ 1 | shown effect ÷ real effect; keep it honest |
| Bars start at zero | truncated bar axes manufacture drama |
| Zoomed line charts | allowed, but flag the non-zero baseline |
That covers the absolute minimum! You can now pick the chart that answers the question, label it so it can be checked, and spot a lying axis at twenty paces — the drawing mechanics are all in TAMYMN-Matplotlib.md.