Task 2: End-to-End ML Pipeline with Scikit-Learn

Objective

Predict customer churn using a production-ready ML pipeline built with Scikit-Learn's Pipeline and ColumnTransformer APIs.

---

Dataset

Telco Customer Churn Dataset
Source: Kaggle / GitHub Mirror

• 7,043 customer records
• 20 features (demographic, service, billing info)
• Target: Churn (Yes/No)

---

Methodology

1. Data Ingestion — Dataset downloaded programmatically via data_ingestion.py
2. Preprocessing Pipeline — Custom transformers for:
   • Cleaning TotalCharges blank strings → NaN → imputed with median
   • Scaling numerical features with StandardScaler
   • Encoding categorical features with OneHotEncoder
3. Model Training — Two models trained with GridSearchCV (5-fold CV):
   • Logistic Regression
   • Random Forest
4. Best Model Export — Saved via joblib to models/best_model.joblib
5. Evaluation — Metrics and plots generated on holdout test set

---

Results

Metric	Score
Accuracy	0.8055
Precision	0.6572
Recall	0.5588
F1 Score	0.6040
ROC-AUC	0.8411

Best Model: Logistic Regression (CV F1: 0.5994 vs Random Forest: 0.5810)

---

Project Structure

├── data/                  # Downloaded dataset
├── models/                # Saved best model (.joblib)
├── notebooks/             # Jupyter Notebook (EDA + full pipeline)
├── reports/               # Confusion Matrix & ROC Curve plots
├── src/
│   ├── data_ingestion.py  # Dataset download
│   ├── pipeline.py        # Preprocessing pipeline
│   ├── train.py           # Training & hyperparameter tuning
│   └── evaluate.py        # Model evaluation
├── requirements.txt
└── README.md

---

How to Run

# 1. Create and activate virtual environment
python -m venv .venv
.venv\Scripts\activate       # Windows

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run pipeline
python src/data_ingestion.py
python src/train.py
python src/evaluate.py

---

Key Observations

• Logistic Regression outperformed Random Forest on this dataset
• ROC-AUC of 0.84 shows strong discriminative ability
• Class imbalance (churn ~26%) caused lower F1 on minority class — expected behavior