Week 9: Advanced & Storytelling Visualizations

🎨 A correct chart and a convincing chart aren't automatically the same thing. This week is about the difference.

🎯 Learning objectives

By the end of this week you can:

• Build faceted/multi-panel visualizations that show a relationship broken down by a third categorical variable.
• Use color, ordering, and annotation deliberately to make a chart's point clear at a glance.
• Choose a colorblind-friendly, purposeful color palette instead of the first default.
• Critique a chart for what it does and doesn't honestly show.
• Save a finished chart to a file for inclusion in a report.

Lesson

Faceting: adding a third dimension with panels

A scatter plot shows two numeric variables; seaborn's faceting tools add a third — usually categorical — dimension by splitting into a grid of small, comparable panels rather than cramming everything into one chart:

import seaborn as sns
import matplotlib.pyplot as plt

sns.relplot(
    data=df,
    x="reading_score",
    y="writing_score",
    col="test_preparation_course",   # one panel per category, side by side
    hue="gender",                     # color within each panel
)
plt.show()

This answers a genuinely multivariate question in one figure: does the reading/writing relationship look different for students who completed test prep vs. not, and does that differ by gender — all at once, comparable side by side because every panel shares the same axis scales. Adding row= alongside col= extends this to a full grid — one dimension across columns, another down rows:

sns.relplot(
    data=df, x="reading_score", y="writing_score",
    col="test_preparation_course", row="lunch", hue="gender",
)

This is now a genuinely 4-variable figure (reading, writing, test prep, lunch, gender — five, actually) — impressive, but worth weighing against the "too many dimensions on one chart" pitfall below.

Ordering categories deliberately

By default, categorical axes are often ordered alphabetically or by first appearance — rarely the most readable order. Explicitly ordering by a meaningful value (like the group's mean) makes a bar chart's pattern immediately legible instead of requiring the reader to hunt for it:

order = df.groupby("parental_level_of_education")["math_score"].mean().sort_values().index

sns.barplot(data=df, x="parental_level_of_education", y="math_score", order=order)
plt.xticks(rotation=45, ha="right")
plt.title("Average math score by parental education level (ordered)")
plt.tight_layout()
plt.show()

sns.barplot (unlike a plain matplotlib bar chart) automatically computes the mean per category and adds error bars representing uncertainty — worth knowing so you don't mistake it for raw counts.

Choosing a deliberate, colorblind-friendly palette

Seaborn's default palette is a reasonable starting point, but purposeful color choices communicate more: use a sequential palette (light to dark) for an ordered numeric variable, a diverging palette (two colors meeting at a neutral midpoint) for values that meaningfully go both above and below some center point (like the correlation heatmap's "coolwarm" in Week 8), and a qualitative palette for unordered categories. Seaborn's "colorblind" palette is a safe, deliberate default whenever color needs to distinguish categories reliably for every reader:

sns.barplot(data=df, x="parental_level_of_education", y="math_score", order=order, palette="colorblind")

Roughly 1 in 12 men have some form of color vision deficiency — defaulting to a colorblind-safe palette costs nothing and excludes no one, which is why it's a reasonable default rather than a special-case choice.

Annotating the point directly

A chart with a specific claim ("students who completed test prep scored, on average, N points higher") is more convincing when that number is visible on the chart itself, not left for the reader to compute from axis positions:

means = df.groupby("test_preparation_course")["math_score"].mean()
ax = means.plot(kind="bar")
for i, value in enumerate(means):
    ax.text(i, value + 1, f"{value:.1f}", ha="center")
plt.ylabel("Average math score")
plt.title("Average math score by test preparation status")
plt.show()

Saving a chart for your report

Once a chart is finished, plt.savefig writes it to an image file — useful for including in a written report rather than only viewing it inline:

plt.savefig("math_score_by_prep.png", dpi=150, bbox_inches="tight")

dpi=150 controls image resolution (higher = sharper, but a larger file); bbox_inches="tight" trims excess whitespace around the figure. Call plt.savefig(...) before plt.show() in the same cell — some environments clear the current figure once it's been displayed.

Honest charts: what to watch for

A few common ways charts mislead, even without any fabricated numbers:

• Truncated y-axis: starting a bar chart's y-axis above 0 exaggerates small differences visually. Bar charts should almost always start at 0; line/scatter plots comparing change are a more defensible exception.
• Cherry-picked axis range: zooming into a narrow x-range can make noise look like a meaningful trend.
• Unlabeled or misleading color scales: a heatmap's color scale should be labeled and, per Week 8, fixed to a consistent range for genuine comparability.
• Overplotting: too many overlapping points (fixed with alpha, per Week 8, or aggregation) can visually hide the actual density of a relationship.

A storytelling chart earns trust by making its honesty easy to verify, not just by looking polished.

⚠️ Common pitfalls

• Stacking too many dimensions onto one chart. row=/col=/hue= together can technically show 4-5 variables at once, but a chart that needs a paragraph of explanation to decode has usually gone past the point of being genuinely more informative than two simpler charts.
• Picking a palette for looks rather than meaning. A rainbow palette on an ordered variable (like exam scores split into 5 ranges) makes the order harder to perceive than a proper sequential palette would.
• Forgetting plt.savefig() must come before plt.show(). Calling them in the wrong order can save a blank image in some environments.
• Confusing sns.barplot's error bars for something else. They represent uncertainty in the estimated mean (by default, a bootstrap confidence interval) — not the data's raw spread, which is what a boxplot (Week 7/8) shows instead.

🧩 Challenges

🧩

Create a faceted scatter plot of math_score vs. reading_score, with one panel per lunch category.

🧩

Build a bar chart of average writing_score by race_ethnicity, with bars explicitly ordered from lowest to highest average.

🧩

Take one of your bar charts from this week and deliberately set plt.ylim(bottom=some_value_above_0). Compare the two versions side by side — how does the truncated version visually mislead about the size of the difference?

🧩

Take a bar chart of group means and annotate each bar with its numeric value directly above it, as shown in the lesson.

🧩

Take one chart from this week that uses color to distinguish categories, and redo it with palette="colorblind". Does the distinction between categories still read clearly?

🧩

Save one of this week's finished charts to a PNG file using plt.savefig, with dpi=150 and bbox_inches="tight".

🤔 Socratic Questions

• Faceting by test_preparation_course and coloring by gender puts four combinations of information on one figure. At what point does adding more dimensions to a single chart start making it harder to read rather than more informative? How would you decide when to split into separate figures instead?
• sns.barplot shows means with error bars, while a plain bar chart of raw counts shows something totally different. What could go wrong if a reader assumed a sns.barplot was showing counts?
• You now know several ways a chart can mislead without any fabricated numbers (truncated axes, cherry-picked ranges, unlabeled scales). Given that, what's your standard for deciding whether a chart you made for your own Week 10 report is "honest," beyond just "the numbers are correct"?
• Sequential, diverging, and qualitative palettes each match a different kind of variable. What would go visually wrong if you used a diverging palette (two colors meeting at a neutral midpoint) on a plain unordered categorical variable like race_ethnicity?
• A colorblind-friendly palette costs nothing to use by default. Can you think of other "accessible by default, costs nothing" choices from earlier weeks (axis labels, chart titles, avoiding color as the only way to distinguish groups) that follow the same principle?

✅ Weekly quiz

✅ Weekly quiz

1. Faceting (e.g. sns.relplot with col=...) is used to:

Remove outliers automaticallyAdd a third (often categorical) dimension via a grid of comparable panelsConvert numeric columns to categoricalCompute correlation automatically

2. Why should a bar chart's y-axis almost always start at 0?

It is a matplotlib requirementStarting above 0 visually exaggerates differences between barsIt makes the chart render fasterIt is only a style preference with no real effect

3. sns.barplot, unlike a plain bar chart of raw values, automatically shows:

Raw counts onlyThe mean per category with error barsA correlation matrixA boxplot instead of bars

4. Explicitly ordering bars by their mean value (rather than alphabetically) mainly improves:

The underlying statisticsReadability — the pattern becomes immediately visibleNothing — order does not affect interpretationThe correlation coefficient

5. A "colorblind" palette is a reasonable default choice because:

It is the only palette seaborn supportsIt ensures category distinctions remain visible to color-vision-deficient readers, at no real costIt only works for numeric dataIt is required by matplotlib

Mark this week complete