M29: From EDA to Narrative: Annotation & Layout
CSCI 1109 — Practical Data Science
❌ Anti-example: when the story breaks
Imagine a dashboard slide:
- Big title: “Our new menu dramatically increased tips”.
- Only figure: bar chart of average tip % before vs after.
- Axes start at 15%, making a small 1% increase look enormous.
- No info on:
- Number of tables in each period.
- Weekday vs weekend mix.
- Lunch vs dinner.
Code
import pandas as pd
import matplotlib.pyplot as plt
# Tiny real difference, but we'll exaggerate it visually
data = pd.DataFrame({
"period": ["Before new menu", "After new menu"],
"avg_tip_pct": [18.2, 19.1] # less than 1 percentage point difference
})
fig, ax = plt.subplots(figsize=(6, 4))
# Bar chart
bars = ax.bar(data["period"], data["avg_tip_pct"])
# Misleading y-axis: start high so the difference looks huge
ax.set_ylim(17, 20) # <- truncation that exaggerates the difference
# Over-confident title
ax.set_title("Our new menu dramatically increased tips", fontsize=14)
ax.set_ylabel("Average tip (%)")
# No mention of sample size, no error bars, no context
# (We could add those in a 'fixed' version, but we deliberately don't.)
# Add value labels on top of bars (helps sell the story)
for bar, val in zip(bars, data["avg_tip_pct"]):
ax.text(
bar.get_x() + bar.get_width() / 2,
val + 0.05,
f"{val:.1f}%",
ha="center",
va="bottom",
fontsize=10
)
# Remove grid to keep the focus on the bars
ax.grid(False)
plt.tight_layout()
plt.show()
Problems:
- Cherry-picking: only the most flattering cut of the data.
- Perception tricks: truncated axis exaggerates differences.
- Missing context: viewer can’t judge reliability.
Anti-example lesson
A clean-looking chart is not automatically honest or informative (Tufte 2001).
Example 1 – EDA scratchpad
Question (vague):
“How does tip % behave as bills get larger?”
Initial EDA pass:
Code
EDA observations
- Tip % is quite spread out, especially for smaller bills.
- Dinner seems to include more large bills than lunch.
- Above some amount (e.g. $40), there are few points.
Example 1 – Turning it into a story figure
Now create a single narrative plot for the restaurant stakeholders.
Design choices
- Audience is now non-technical staff at the restaurant.
- Title states the main message in plain language.
- Shaded band marks a “normal” tipping range (15–20%).
- Axes are trimmed to the region with most data, but y still starts at 0 to avoid exaggeration.
- A single annotation calls out unusually generous tips on mid-sized dinner bills.
- Many EDA variants → one clear, annotated figure.
Code
fig, ax = plt.subplots(figsize=(7, 4))
sns.scatterplot(
data=tips,
x="total_bill",
y="tip_pct",
hue="time",
alpha=0.5,
ax=ax,
legend=True,
)
# Focus on the main range where most customers live
ax.set_xlim(0, 60)
ax.set_ylim(0, 40)
ax.set_title("Most customers tip around 15–20% on typical dinner bills")
ax.set_xlabel("Total bill ($)")
ax.set_ylabel("Tip (% of bill)")
# Add a horizontal band for 15–20% as a “normal tipping” region
ax.axhspan(15, 20, color="grey", alpha=0.1)
ax.text(
5, 21.5,
"Common tipping range\n(15–20%)",
fontsize=9,
va="bottom"
)
# Optional annotation: call out high tips on mid-sized dinner bills
ax.annotate(
"Some dinner parties tip 25%+ on $30–$50 bills",
xy=(40, 26),
xytext=(48, 32),
arrowprops=dict(arrowstyle="->"),
fontsize=9,
)
ax.legend(title="Time of day", loc="lower right")
plt.tight_layout()
Example 2 – Narrative bar chart with labels
Code
fig, ax = plt.subplots(figsize=(7, 4))
sns.barplot(
data=summary_day_time,
x="day",
y="mean_pct",
hue="time",
ax=ax,
)
ax.set_ylabel("Average tip (%)")
ax.set_xlabel("Day of week")
ax.set_title("Weekend dinners bring the highest tip percentages")
# Add labels on bars for quick reading
for container in ax.containers:
ax.bar_label(container, fmt="%.1f", padding=2, fontsize=8)
ax.legend(title="Time of day", loc="lower right")
plt.tight_layout()
Design choices
- Single, centered figure instead of conflicting mini-charts.
- Ordering of days (Thur → Sun) makes weekend pattern easier to see.
- Bar labels let viewers get values without guessing from the y-axis.
- Title states the main takeaway; legend reinforces colour coding.
