Telco Customer Churn Prediction

Project Overview

This project develops a high-sensitivity machine learning pipeline to identify customers at risk of churn for a telecommunications provider. By shifting focus from standard accuracy to high-recall modeling, the system ensures that approximately 82% of actual churners are flagged for proactive retention efforts.

Project Structure

The repository is organized into a modular architecture to ensure scalability and maintainability:

• data/: Contains raw and processed datasets.
• notebooks/: Jupyter notebooks for EDA, experimental modeling, and extracting actionable business insights to reduce customer attrition.
• src/: Core logic including data preprocessing, visualization, and model training scripts.
• requirements.txt: List of dependencies required to run the environment.

Methodology

1. Exploratory Data Analysis (EDA)

• Identified significant class imbalance (approx. 26% churn rate).
• Analyzed key correlations between tenure, Contract, InternetService, and churn probability.

2. Modeling & Strategy

• Baseline: Logistic Regression used to establish a linear performance floor.
• Candidate Models: Random Forest and XGBoost selected for their ability to handle non-linear relationships.
• Optimization: Conducted exhaustive Hyperparameter Tuning via GridSearchCV.
• Imbalance Handling: Utilized scale_pos_weight and class_weight to prioritize Recall over standard Accuracy.

3. Model Interpretation (SHAP)

Utilized SHapley Additive exPlanations to provide transparency into model decisions. Key churn drivers include:

• Contract Type: Month-to-month contracts are the highest risk factor.
• Tenure: Lower tenure strongly correlates with increased churn risk.
• Service Tier: Fiber Optic users exhibit higher attrition rates.

Results

Model	Recall (Churn)	F1-Score	AUC-ROC
Optimized XGBoost	0.82	0.62	0.844
Optimized Random Forest	0.76	0.63	0.844
Logistic Regression	0.55	0.59	0.839

Installation & Usage

1. Clone the repository.
2. Install dependencies: pip install -r requirements.txt.
3. Run the analysis: Execute the main notebook or training scripts in src/.