Customer Churn Classifier
Part of the AI Engineer Roadmap
In short
Winning a new customer usually costs more than keeping an old one. A churn model flags who is about to leave, so the business can act before they go.
A binary classification project: predict whether a customer will churn — a business-relevant task that forces you to handle imbalanced classes and communicate results.
Think of it like this
The recharge shop uncle
The uncle at the mobile recharge shop knows his regulars. When someone who came every month hasn't shown up in weeks and complained about the network last time, he guesses they've switched SIMs. A churn model makes that same guess, for lakhs of customers.
Watch it happen
Your build plan
1/6Use the public Telco Customer Churn dataset. Each row is a customer, Churn is Yes or No.
The big ideas
What you need to know
Classes are imbalanced
If 75% of customers stay, a model that says 'everyone stays' scores 75% accuracy and catches zero leavers.
Recall on churners matters
Recall answers: of the people who left, how many did we flag? That's usually what the business cares about.
Tune the threshold
predict_proba gives a churn probability. Lowering the cutoff from 0.5 catches more leavers at the cost of more false alarms.
Explain the why
Managers act on reasons, not scores. Show which features push churn risk up, using coefficients or feature importances.
See it in code
Walk through the code
1import pandas as pd2from sklearn.linear_model import LogisticRegression3from sklearn.metrics import classification_report4from sklearn.model_selection import train_test_split5 6df = pd.read_csv("telco_churn.csv")7y = (df["Churn"] == "Yes").astype(int)8X = pd.get_dummies(df[["tenure", "MonthlyCharges", "Contract"]], drop_first=True)9 10X_tr, X_te, y_tr, y_te = train_test_split(X, y, stratify=y, random_state=1)11model = LogisticRegression(class_weight="balanced", max_iter=1000)12model.fit(X_tr, y_tr)13print(classification_report(y_te, model.predict(X_te)))1/4Turn Yes/No into 1/0, and expand Contract into 0/1 columns with get_dummies.
class_weight='balanced' makes each missed churner cost more during training, so the model stops ignoring the smaller class.
Save yourself the pain
Common mistakes
Oops
Celebrating 80% accuracy without checking how many churners were caught.
Do this instead
Look at recall for the churn class and the confusion matrix. Accuracy hides the minority class.
Oops
Keeping columns like 'cancellation_date' that only exist after a customer leaves.
Do this instead
Use only what you'd know before the churn happens. Future information is leakage and fakes a great score.
Your mission
Milestone 1: a fair first model
Quick check
Your churn model has 85% accuracy but only 20% recall on churners. What does that mean?
Go deeper
- 1Use a real churn dataset and handle class imbalance explicitly.
- 2Report precision/recall trade-offs, not just accuracy, since churn is usually imbalanced.
- 3Translate model output into a business recommendation (e.g. who to target for retention).
More projects at this stop
House Price Prediction
Stuck on this topic? Ask an Insider
Get 1:1 guidance from people who've walked this exact path — on the InsideEdge app.
