Telecommunications Customer Churn Prediction using Machine Learning

Project Overview:

An end-to-end real-world project developed for the telecom operator Djezzy, aimed at predicting customer churn to enable proactive retention strategies.

Technical Challenges Solved:

Handled a massive and complex dataset containing over 8.5 million users.

Successfully tackled severe Class Imbalance, where churn cases represented only 1% of the total dataset.

Methodology & Techniques:

Performed extensive exploratory data analysis (EDA), data preprocessing, and feature engineering to extract the most impactful variables.

Addressed the extreme data imbalance using advanced techniques such as Undersampling and Cost-Sensitive Learning.

Trained, evaluated, and fine-tuned state-of-the-art machine learning algorithms: XGBoost, LightGBM, Random Forest, and Logistic Regression.

Achieved high predictive performance with an AUC of 0.93 and a Recall of 0.90 using LightGBM combined with Cost Learning.

Implemented Uplift Modeling to isolate the true positive anomalies and target the most receptive customers, significantly reducing False Positives and optimizing marketing resources.

Final Product:

Developed and deployed an interactive, user-friendly web dashboard using Streamlit, allowing the sales and marketing teams to visualize statistics, analyze predictions, and easily export actionable insights.