Enterprise Telecom Churn Prediction: A Value-Driven ML Pipeline

Objective: Bridging the gap between raw mathematical accuracy and real-world business retention strategy.

⬢ Executive Summary

Customer churn costs enterprise telecommunications companies millions annually. Standard machine learning models often optimize for global accuracy, completely missing the minority class (the actual churning customers).

This project discards the baseline approach. By implementing synthetic data generation (SMOTE), hardware-accelerated ensemble learning (XGBoost via CUDA), and strategic probability threshold calibration, this pipeline successfully intercepts 96% of churning customers—translating predictive analytics into direct revenue retention.

⬢ Technical Architecture

This is a full-stack, end-to-end data pipeline optimized for a local NVIDIA GPU environment:

1. Data Ingestion & Formatting: - Handled hidden data corruption (e.g., coerced whitespace errors in continuous numerical columns) and enforced strict float64 and Int64 typing.
   • Dropped non-predictive cardinality (Customer IDs) to prevent model overfitting.
2. Feature Engineering:
   • Translated human business context into algorithm-ready mathematics via robust One-Hot Encoding (pd.get_dummies) and Boolean to Integer conversion.
   • Applied StandardScaler to ensure uniform loss-landscape navigation during gradient descent.
3. Class Imbalance Resolution:
   • Implemented SMOTE (Synthetic Minority Over-sampling Technique) strictly on the X_train matrix, mathematically generating realistic minority vectors to force the algorithm to learn churn patterns rather than memorizing class biases.
4. Hardware-Accelerated Modeling (XGBoost):
   • Bypassed CPU bottlenecks by routing the XGBoost hist tree method directly to a dedicated NVIDIA GPU (device='cuda').
   • Deployed GridSearchCV with 3-fold cross-validation to isolate the optimal hyperparameter architecture (Max Depth: 3, Learning Rate: 0.01) specifically targeting the Recall metric.

⬢ Business Value & Impact

A model is only as valuable as the money it saves. By aggressively lowering the prediction threshold from the default 0.50 to a customized 0.35, the model was tuned to reflect the reality of B2C marketing: sending a $10 retention discount to a false positive is infinitely cheaper than losing a $1,000 customer entirely.

Metric	Baseline Model (Logistic Regression)	Tuned Architecture (XGBoost + SMOTE)	Business Impact
Recall (Churn)	60%	96%	Intercepts nearly 100% of flight-risk accounts.
Precision (Churn)	68%	38%	Widens the net. Accepts controlled false alarms to guarantee maximum retention.

⬢ Local Deployment

To replicate this pipeline locally, ensure your environment is configured for GPU acceleration.

# Clone the repository
git clone [https://github.com/jemmziray-tech/enterprise-churn-interception.git](https://github.com/jemmziray-tech/enterprise-churn-interception.git)

Install strictly defined dependencies

pip install -r requirements.txt

Note: Ensure local NVIDIA drivers are optimized for CUDA acceleration to utilize the XGBoost GPU tree method.

Engineered by John Elifuraha Mziray — BSc Artificial Intelligence & Machine Learning, University of Limerick.