Skip to content

Exercises — Week 4 — Charts That Change a Decision

Helen's board meeting is Monday. She asked for one screen, not a notebook. Rebuild the deck's plan-churn bar honestly, and give her the adoption and region cuts she'll ask about when someone pushes back.

What you are building

Three charts: adoption by plan, churn by region, and an honest plan-churn bar whose title is a claim and whose y-axis starts at 0.

Predict before you run

  1. Which plan adopts the most features?
  2. If the y-axis starts at 14% instead of 0, what lie does the screenshot tell?
  3. Should region churn be computed on events or on users?

Before you start

  • Use the Agg backend (matplotlib.use("Agg")) in a terminal so matplotlib writes PNGs instead of blocking on a window.

Each task has three hints, closed by default. Open only as far as you need.

Task

Work in starter.py. Run from the repo root:

python exercises/ml/week-04/starter.py

1. Adoption curve. For users with a signup_date, plot average features_adopted by plan_type as a bar. Annotate the winner.

Hint 1 — a nudge

A chart is an API response. If Helen has to squint to find the winner, the response is incomplete — put the number on the bar.

Hint 2 — the approach

load_customer_360() already has features_adopted and signup_date. groupby("plan_type")["features_adopted"].mean(), ax.bar, then ax.annotate the idxmax() with its value.

Hint 3 — most of the code 
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
import pandas as pd

from lib.course_data import find_data_dir, load_customer_360

DATA = find_data_dir()
df = load_customer_360(DATA)
adopt = df[df["signup_date"].notna()].groupby("plan_type")["features_adopted"].mean().sort_values()

fig, ax = plt.subplots(figsize=(6, 3.5))
ax.bar(adopt.index, adopt.values, color="#6366f1")
top = adopt.idxmax()
ax.annotate(f"{adopt[top]:.2f}", (top, adopt[top]), ha="center", va="bottom")
ax.set_title("Average features adopted by plan")  # rewrite as a claim
fig.savefig("adoption_by_plan.png", dpi=120)

2. Region bars. Most-common region per user from events, then churn rate by region. Horizontal bars, sorted.

Hint 1 — a nudge

Should a region with chatty users count more than a region with quiet ones? Decide what one "vote" is before you group.

Hint 2 — the approach

Same collapse as Week 2: drop null regions, one mode per user_id, left-join onto the 360. Only then groupby("region")["is_churned"].mean(), sort_values(), ax.barh.

Hint 3 — most of the code 
events = pd.read_csv(DATA / "user_events.csv", usecols=["user_id", "region"]).dropna()
region = events.groupby("user_id")["region"].agg(lambda s: s.mode().iat[0]).rename("region")
by_user = df.merge(region, on="user_id", how="left")
churn = by_user.groupby(by_user["region"].fillna("unknown"))["is_churned"].mean().sort_values()

fig, ax = plt.subplots(figsize=(6, 3.5))
ax.barh(churn.index, churn.values, color="#0f766e")
ax.set_xlabel("churn rate (users)")
fig.savefig("churn_by_region.png", dpi=120)

3. Honest title. Rebuild the plan-churn bar so the title is a claim ("Free churn is ~2× paid") and the y-axis starts at 0.

Hint 1 — a nudge

Screenshot the chart and cover the axis labels. What would a board member conclude from the bar heights alone?

Hint 2 — the approach

Compute churn by plan, ax.bar, then ax.set_ylim(0, ...) explicitly — matplotlib autoscaling is not a promise. The title is a sentence someone could disagree with.

Hint 3 — most of the code 
plan_churn = df.groupby("plan_type")["is_churned"].mean().sort_values(ascending=False)

fig, ax = plt.subplots(figsize=(6, 3.5))
ax.bar(plan_churn.index, plan_churn.values, color="#b91c1c")
ax.set_ylim(0, plan_churn.max() * 1.15)
ax.set_ylabel("churn rate")
ax.set_title("TODO: one sentence the numbers support")
fig.savefig("plan_churn_honest.png", dpi=120)

Success criteria

  • Winner annotated on the adoption bars.
  • Region chart is user-grain.
  • Honest chart: title is a sentence, ylim starts at 0.

After you run

A chart is an API response. If the PM can misquote the title, rewrite the title. This is the screen Helen actually shows the board.

Week 4 — Charts That Change a Decision