Telecom Customer Churn Prediction

Live link - https://web-app-churn-prediction-1.onrender.com/

A machine learning project to predict customer churn in the telecommunications industry using various classification algorithms.

Overview

This project analyzes customer data to identify patterns that lead to customer churn and builds predictive models to help telecom companies retain customers proactively.

Features

• Data Analysis: Comprehensive EDA with correlation heatmaps, churn rate analysis by demographics
• Multiple Algorithms: Logistic Regression, Random Forest, XGBoost, CatBoost, and more
• Model Optimization: Hyperparameter tuning and threshold optimization for improved precision
• Visualization: ROC curves, Precision-Recall curves, and performance metrics
• Feature Engineering: Missing value handling and categorical encoding

Dataset Features

• Demographics: Age, Gender, Income
• Usage: Data usage (GB), Tenure (months)
• Service: Plan type (Prepaid/Postpaid), Number of complaints
• Target: Customer churn (0 = retained, 1 = churned)

Key Insights

• Plan Type Impact: Prepaid customers show 28% churn rate vs 20% for Postpaid
• Complaints Correlation: Strong positive correlation between complaints and churn
• Tenure Effect: Longer tenure customers are less likely to churn
• Model Performance: Achieved 78.2% AUC with optimized Logistic Regression

Model Performance

• Best Model: Tuned Logistic Regression
• AUC Score: 0.782
• Average Precision: 0.552
• Key Features: Complaints, tenure, plan type, usage patterns

Files

• BEST_MODEL.ipynb: Main end-to-end notebook (training, tuning, export artifacts)
• datasets/churn_data.csv: Generated synthetic dataset (10,000 records)
• README.md: Project documentation

Technologies Used

• Python: pandas, numpy, scikit-learn
• ML Libraries: XGBoost, CatBoost, scikit-learn
• Visualization: matplotlib, seaborn
• Model Evaluation: ROC curves, Precision-Recall analysis

Streamlit App

A ready-to-run Streamlit app is provided to interactively predict churn for a single customer.

Prerequisites

• Python 3.10+
• Recommended: a virtual environment

Install Dependencies

pip install -r requirements.txt

Export Model Artifacts (if not already present)

Run the BEST_MODEL.ipynb to generate and save artifacts into deployment/ (preferred) or models/ directories. The app will automatically pick the latest artifacts.

Expected files:

• deployment/best_model_*.pkl (or models/best_*.pkl)
• deployment/preprocessor_*.pkl (or models/enhanced_preprocessor.pkl)
• deployment/model_metadata_*.pkl (optional)

Run the App

streamlit run app.py

Open the URL shown in the terminal (typically http://localhost:8501).

Notes

• The app recreates the same feature engineering used in training to ensure consistency.
• If a full pipeline was saved, it is used directly. Otherwise, the app loads the preprocessor and applies it before prediction.
• Artifacts are auto-discovered from deployment/ first, then models/ as a fallback.

Business Impact

This model can help telecom companies:

• Identify high-risk customers for targeted retention campaigns
• Optimize customer service resources
• Reduce churn rates and improve customer lifetime value
• Make data-driven decisions about customer retention strategies

Future Enhancements

• Real-time prediction API
• Customer segmentation analysis
• A/B testing framework for retention strategies
• Integration with CRM systems