DATAVIZ 101: Charts that answer questions

Eight lessons on the last step of the pipeline, where the numbers you cleaned and grouped in PANDAS 101 become something a person takes in at a glance. Every chart in this course is drawn from correct data. Several of them mislead anyway, and taking them apart is the course.

How this course teaches

Same tutor as the rest of the series: predict before you run, and hints that climb a ladder ending in an explanation rather than pasted code. What is different here is that a chart cannot fail on its own. Plotting code succeeds whether or not the picture is honest, so every graded cell reads its own figure back and checks it:

assert ax.get_ylim()[0] == 0, "a bar is read by its length"
assert ax.collections[0].get_array() is not None, "the third column is missing"
assert np.diff(ax.lines[0].get_xydata()[:, 0]).max() <= 1, "that segment crosses a gap"

An axes is an object with a readable state, which is what turns "does this chart lie" from an opinion into a check. It is also the most useful habit the course can leave you with, because the same three lines work on a figure somebody else wrote.

What you will be able to do

• Build a figure whose parts you can name and read back: figure, axes, title, labels, limits, legend
• Choose between a line, a cloud of points and bars by what the horizontal axis means, and catch a line that crosses a gap in the same ink it uses for data
• Pick a bin count you can defend with a split half check, and say what fell outside the range
• Read a box plot as five numbers, and spot a box whose height is a mixture of two populations rather than a spread
• Put a third column on a scatter, and show a confound instead of arguing about one
• Replace a grouping loop with one seaborn call, and name what that call decided for you
• Measure how much a chart exaggerates, as a number, and undo it without changing any data

The lessons

The figure

1. Anatomy of a figure
2. Lines and scatters

Distributions 3. Histograms and bins 4. Box plots and spread

Relationships 5. Scatter and trend 6. Seaborn in one move

Judgment 7. Charts that lie 8. Pick and defend

The lot

360 used car listings, written to disk by each lesson's setup cell from a fixed seed, so any lesson opens on its own and rerunning is always safe. It is the PANDAS 101 forecourt a season on, including the 1971 Kessel Adler that dragged the mean around there and is still unsold.

Every property of it feeds a lesson. Prices are two populations eleven thousand lira apart, which is a valley in lesson 3 and a false spread in lesson 4. Three classics ask 139,500 and up, which is what hands the whole axis to an empty region in lesson 3 and puts three dots past the whisker in lesson 4. Kilometres accumulate with the model year, so lesson 5 gets a correlation that is minus 0.11 across the lot, minus 0.84 inside one body and minus 0.03 inside one model year. Two brands ask within 2.2 percent of each other, which lesson 7 turns into a chart where one bar looks two and a half times the other.

Lesson 8 removes the file.

Where this course sits

Fifth course of the ivyx AI series and the second half of the data wave, after PANDAS 101, which is a real prerequisite: every chart here is drawn from a groupby, a filter or a merge you can already write. The judgment it installs carries forward more than the syntax does. ML 101 opens with a baseline and a scatter of predictions against truth, and every model from there on reports itself as a loss curve, which is a line over an ordered axis whose slope is the whole story. PYTORCH 102 spends a whole lesson reading those curves. You will arrive knowing what to check before believing one.

Requirements

Python 3, with pandas, numpy, matplotlib and seaborn. Verified against matplotlib 3.11 and seaborn 0.13. Two keywords in the course need recent versions: tick_labels on ax.boxplot needs matplotlib 3.9 or newer, and errorbar on sns.barplot needs seaborn 0.12 or newer.