Prediction of lost customers in the telecommunications industry utilizing machine learning on large data platforms

Abstract

The telecom industry study is crucial for boosting businesses' profitability, particularly by
accurately predicting churn. This research focused on developing a customized churn prediction
system for SyriaTel, a telecom company. High AUC values were essential for precise churn
forecasts, and the dataset was split into 70% training and 30% testing sets. Cross-validation
aided in reliable model assessment and hyper parameter tuning. Feature engineering and
selection techniques were employed to prepare the features for machine learning algorithms.
Addressing data imbalance, under-sampling and tree-based algorithms were utilized. Four treebased
models were chosen: Decision Tree, Random Forest, Gradient Boosting Machine, and
XGBOOST. Success relied on strategic planning and inclusion of mobile social network features.
XGBOOST outperformed with a 93.301% AUC on the SyriaTel dataset, followed by GBM,
Random Forest, and Decision Tree. Testing with a new dataset showed XGBOOST's AUC at 89%.
Regular model retraining is necessary due to non-stationary data. Incorporating Social Network
Analysis improved churn prediction in telecom