Predict customer churn on a synthetic dataset using Python. The pipeline includes data generation, feature engineering, model training (Logistic Regression, Random Forest, Gradient Boosting), hyperparameter search, class weighting, selection by PR-AUC, and decision-threshold tuning to balance precision and recall. Outputs metrics, reports, and visualizations.

• Synthetic customer dataset with realistic behavior signals
• Models: Logistic Regression, Random Forest, Gradient Boosting
• Hyperparameter optimization (RandomizedSearchCV) & class weighting
• Model selection by PR-AUC (Average Precision)
• Threshold tuning (F2 focus) with precision floor
• Metrics: Accuracy, Precision, Recall, F1, ROC-AUC, PR-AUC
• Visuals: ROC, Precision-Recall, Confusion Matrix, Feature Importance
• Saved artifacts: best model (joblib) & metrics

customer-churn-prediction/
├─ README.md
├─ LICENSE
├─ requirements.txt
├─ data/
│  └─ generate_customers.py
├─ src/
│  ├─ train_models.py
│  └─ utils.py
└─ outputs/
   └─ figures & reports

python -m venv .venv
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate
pip install -r requirements.txt

python data/generate_customers.py --n 10000 --seed 42 --out data/customers.csv

python src/train_models.py --input data/customers.csv --outdir outputs --test-size 0.2 --val-size 0.2 --seed 42

Outputs

• outputs/metrics.json – model choice, tuned threshold, test metrics
• outputs/classification_report.txt
• outputs/roc_curve.png
• outputs/pr_curve.png
• outputs/confusion_matrix.png
• outputs/feature_importance.png
• outputs/best_model.joblib

➡️ The model now catches ~50% of churners with precision ~0.52, balancing false positives and recall.