Project — Customer Churn Classification

Relax. We'll talk through this in plain words — no textbook voice.

What you'll walk away with

Understand Project — Customer Churn Classification without any of the intimidation
Be able to run scikit-learn code yourself
Apply this concept in a real project right away

Let's stop and think about this for a second

Customer Churn Prediction is one of the most common classification use cases in business — using a customer's behavior data (usage frequency, support ticket count, contract length) to predict whether they're likely to cancel their subscription in the coming weeks, so the business team can proactively send a retention offer to high-risk customers. Class imbalance (churned customers being far fewer than non-churned ones, as is typical in the real world) shows up in this project, which means you'll be putting the Intermediate chapter's Precision/Recall/F1 concepts directly to use.

Let's connect it to a real scenario

Take a customer dataset (usage_frequency, support_tickets, contract_length, churned) and train both a Logistic Regression and a Random Forest Classifier, compare Precision/Recall/F1 with `classification_report()`, then check the Random Forest's `feature_importances_` and write up a short report to the business team on which factors best predict churn.

Customer Churn Prediction = a business-relevant classification use case
Properly evaluated class imbalance using Precision/Recall/F1 (not just Accuracy alone)
Pulled out a business insight (which factor drives churn most) using feature_importances_

Let's look at it together

python

from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report

model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)

predictions = model.predict(X_test)
print(classification_report(y_test, predictions))

# Which features matter most for predicting churn?
importances = sorted(zip(X.columns, model.feature_importances_), key=lambda x: -x[1])
for feature, importance in importances:
    print(f"{feature}: {importance:.2f}")

You should see

              precision    recall  f1-score
           0       0.91      0.95      0.93
           1       0.78      0.65      0.71

support_tickets: 0.42
usage_frequency: 0.35
contract_length: 0.23

Try it in 5 minutes

Create your own customer churn sample dataset, run it through the full classification pipeline, then look at the `feature_importances_` results and write a one-paragraph insight for the business team.

One thing to watch out for

When using customer data for an ML model, you need to follow privacy/data protection regulations (e.g. GDPR) — using real customer data for model training without authorization or consent can create legal problems.

Easy traps

Seeing a single 'accuracy 91%' number and jumping straight to 'the model's good' without checking the Recall on the churn class (the minority class), which might be sitting at just 65% — always check the minority class's metrics carefully
Directly assuming a feature importance result means causation ('this factor causes churn') — it only shows correlation, and any business decision built on it needs a domain expert's confirmation first

Now try it yourself

You'll know it worked when: precision recall f1-score 0 0.91 0.95 0.93 1 0.78 0.65 0.71 support_tickets: 0.42 usage_frequency: 0.35 contract_length: 0.23