End-to-end telecom churn prediction system using machine learning. Includes EDA, feature engineering, RFE/PCA-based modeling, and recall-optimized evaluation. Identifies at-risk customers using behavioral trends and enables data-driven retention strategies for improved customer lifetime value.
This project builds a complete end-to-end machine learning system to predict customer churn in the telecom industry.
The goal is to identify high-risk customers early and enable data-driven retention strategies.
Customer churn leads to significant revenue loss.
This project answers:
- Who is likely to churn?
- Why are they churning?
- How can we intervene early?
- 📊 Exploratory Data Analysis (EDA)
- 🧹 Data Cleaning & Feature Engineering
- 🔍 Feature Selection (RFE, Correlation, VIF)
- ⚙️ Dimensionality Reduction (PCA)
- 🤖 Model Building:
- Logistic Regression
- Random Forest
- Gradient Boosting
- XGBoost
- 🎯 Threshold Optimization (Recall-focused)
- 📈 Model Evaluation (ROC, PR Curve, Confusion Matrix)
├── data/
│ ├── train.csv
│ └── test.csv
├── notebooks/
│ └── Artificial-Intelligence-Driven-Customer-Retention-System.ipynb
├── README.md
└── requirements.txt
- Python 🐍
- Pandas, NumPy
- Scikit-learn
- Statsmodels
- XGBoost
- Matplotlib, Seaborn
- 📉 ARPU trends (revenue decline)
- 📞 Call usage patterns
- ⏱ Recharge gap features
- 📅 Temporal behavior changes
- ⏳ Customer tenure
| Model | Accuracy | Recall (Churn) |
|---|---|---|
| Logistic Regression | ~75% | ~82% ✅ |
| PCA + Logistic | ~76% | ~82% 🔥 |
| Gradient Boosting | ~92% | ~23% ❌ |
| XGBoost | ~92% | ~35% ❌ |
- High Recall (~82%)
- Stable Generalization
- Handles Multicollinearity
- Business-aligned performance
- 📉 Declining revenue is the strongest churn signal
- 📞 Reduced usage indicates disengagement
- ⏱ Recharge delays are early churn indicators
- 📅 Recent behavior matters more than historical
- 🔄 Churn is a gradual behavioral process
- 🎯 Segment users by churn risk
- 🔴 High risk → aggressive retention
- 🟠 Medium risk → engagement campaigns
- 🟢 Low risk → no action
Raw Data → Cleaning → Feature Engineering → Scaling → PCA → Model → Prediction
# Install dependencies
pip install -r requirements.txt
# Run notebook
jupyter notebook- Deploy using FastAPI / Streamlit
- Real-time churn prediction system
- Advanced models (LightGBM, tuned XGBoost)
- Cost-sensitive learning
- Churn is not a sudden event — it is a gradual disengagement process.
Feel free to fork and improve the project!
For any queries or collaboration, reach out!
⭐ If you like this project, give it a star!