Machine Learning2024github.com/kmexa

Customer Churn Prediction: Production-Ready ML Pipeline

Production-grade churn prediction pipeline demonstrating zero-data-leakage architecture, reproducible preprocessing, and evaluation tied directly to business outcomes. Gradient Boosting tuned with GridSearchCV achieves ROC-AUC 0.934 and PR-AUC 0.812.

Download Notebook (.ipynb) View on GitHub

Methodology

01Feature engineering: AvgMonthlyCharges = TotalCharges / tenure captures customer value trajectory across the contract lifetime

02ColumnTransformer: StandardScaler for numerical features, OneHotEncoder for categorical - all inside sklearn Pipeline to guarantee zero data leakage across CV folds

03GridSearchCV over n_estimators x max_depth x learning_rate grid with stratified 5-fold cross-validation

04Threshold optimisation: default 0.5 threshold is rarely optimal; F1-maximising threshold found on validation set to balance precision-recall for the retention use case

05Feature importance mapped to concrete retention strategies with business annotation

06Business impact analysis: cost of false negatives (missed churners) versus false positives (unnecessary retention spend) quantified

Results

0.934

ROC-AUC

0.812

PR-AUC

7,043

Records

Model	ROC-AUC	PR-AUC
Logistic Regression	0.851	0.669
Random Forest	0.912	0.773
Gradient Boosting (tuned)	0.934	0.812

Gradient BoostingColumnTransformerPipelineGridSearchCVscikit-learnpandasPython

Download customer_churn.ipynb Request Similar Project

Customer Churn Prediction: Production-Ready ML Pipeline

Methodology

Results

More Portfolio Projects