Customer Churn Prediction - End-to-End ML Pipeline

Click here to watch tutorial

📋 Project Overview

A production-ready machine learning pipeline to predict customer churn in the telecommunications industry. This project demonstrates end-to-end ML workflow including data preprocessing, feature engineering, model training, evaluation, and deployment-ready model artifacts.

Business Problem: Identifying customers likely to churn helps businesses take proactive retention measures, reducing revenue loss.

🎯 Key Results

• Best Model: Logistic Regression
• ROC-AUC Score: 0.8458
• Accuracy: 80.41%
• F1-Score: 0.5929

📊 Dataset

• Source: Telco Customer Churn Dataset
• Size: 7,043 customers
• Features: 20 (demographic, account, and service information)
• Target: Churn (Yes/No)
• Churn Rate: ~26.5%

Features Include:

• Demographics: Gender, Senior Citizen, Partner, Dependents
• Account Info: Tenure, Contract type, Payment method
• Services: Phone, Internet, Online Security, Tech Support, etc.
• Charges: Monthly charges, Total charges

🛠️ Tech Stack

• Language: Python 3.8+
• ML Libraries: scikit-learn, XGBoost, imbalanced-learn
• Data Processing: Pandas, NumPy
• Visualization: Matplotlib, Seaborn
• Model Persistence: Joblib

🚀 Getting Started

Prerequisites

python >= 3.8
pip

Installation

1. Clone the repository

git clone https://github.com/misrapk/churn-prediction-ml.git
cd churn-prediction-ml

2. Create virtual environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Download dataset from Kaggle and place in data/ folder

Usage

1. Run EDA:

jupyter notebook notebooks/01_eda.ipynb

2. Train Models:

jupyter notebook notebooks/03model_training.ipynb

3. Use Preprocessor (standalone):

python src/data_preprocessing.py

🔍 Methodology

1. Data Preprocessing

• Handled missing values in TotalCharges
• Created engineered features:
  • tenure_group: Categorized tenure into bins
  • avg_monthly_per_tenure: Average spending rate
  • num_services: Count of subscribed services
• Encoded categorical variables (Label Encoding & Binary Encoding)
• Scaled numerical features using StandardScaler

2. Model Training

Trained and compared 5 models:

• Logistic Regression (baseline)
• Decision Tree
• Random Forest
• Gradient Boosting
• XGBoost

3. Evaluation Metrics

• Accuracy
• Precision
• Recall
• F1-Score
• ROC-AUC (primary metric for imbalanced data)

4. Model Selection

Selected [Your best model] based on highest ROC-AUC score, balancing precision and recall for business needs.

Key Insights

1. Contract Type is the strongest predictor - month-to-month contracts have 3x higher churn
2. Tenure inversely correlates with churn - customers with <12 months tenure churn most
3. Monthly Charges - higher charges correlate with increased churn
4. Tech Support subscription reduces churn by 40%

Visualizations

🔮 Future Improvements

• Implement hyperparameter tuning (GridSearchCV/RandomizedSearchCV)
• Handle class imbalance with SMOTE/undersampling
• Build REST API with FastAPI for model serving
• Add CI/CD pipeline for automated retraining
• Deploy on AWS/Azure as web service
• Create Streamlit dashboard for predictions

📝 Lessons Learned

1. Feature engineering significantly improved model performance (+8% ROC-AUC)
2. Class imbalance handling crucial for churn prediction
3. Tree-based models outperform linear models for this problem
4. Business context matters - optimizing for recall may be more valuable than accuracy

👤 Author

Peeyush K Misra

• LinkedIn: peeyushkmisra
• GitHub: @misrapk
• Email: peeushkmisra@gmail.com

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Dataset provided by IBM Watson Analytics
• Inspired by real-world telecom churn challenges