Leveraging TabNet for Enhanced Customer Churn Prediction in the Telecommunication Industry

In the telecommunications sector, keeping existing customers is crucial because the costs associated with acquiring new customers are substantially higher than those for retaining existing ones. To address this, our study employs TabNet, a deep learning model that employs sequential attention for decision-making in tabular data, to enhance the prediction of customer churn using IBM Telco Dataset. We applied various data balancing methods to address the issue of imbalanced class distributions. These techniques included the oversampling method, Synthetic Minority Over-sampling Technique (SMOTE), as well as hybrid approaches that combine oversampling and undersampling, specifically SMOTE Edited Nearest Neighbors (SMOTEENN) and SMOTE Tomek Links (SMOTETomek). Our methodology is validated via a 10-fold cross-validation and achieved more recognizability than the state-of-the-art with a 96.60% accuracy and a 96.91 F1-score.