Explainable AI-Driven Triple Adaptive Feature Selection Model for Enhanced Telecom Churn Prediction

Customer churn is still a very important issue in the competitive telecommunication industry, as it has a direct impact on the profitability and sustainability of the organization. Although machine learning and deep learning models have achieved a lot of success in customer churn prediction tasks, they are often less interpretable and may not generalize well to different datasets. This work presents an explainable and adaptive hybridized churn prediction framework that combines data preprocessing, feedback-driven intelligent feature selection, class rebalancing, and explainability in a single modular architecture. An adaptive feature selection controller is introduced to dynamically refine the feature subset through iterative feedback based on feature stability and redundancy reduction to provide reduced dimensionality. Mutual Information (MI), Recursive Feature Elimination (RFE), and Boruta are integrated within an adaptive hybrid feature selection mechanism that continuously updates the selected feature space rather than performing static preprocessing. Furthermore, statistical significance testing is conducted to prove the effectiveness of the adaptive feature selection model over individual feature selection methods. To counter the class imbalance problem in telecom data, Random Minority Oversampling is used to improve the representation of minority classes. The optimized features are then used to train a combination of predictive models, such as ML and DL classifiers, and a transformer model to ensure a balance between accuracy and explainability. Additionally, the inclusion of Explainable AI (XAI) modules such as SHAP and LIME enables the interpretability of the model, enabling stakeholders to understand the reasoning behind churn predictions. Experimental results were conducted on two benchmark datasets, including Cell2Cell and IBM Telco, which provided an accuracy of 88.16%, 89.60%, and an F1 score of 0.878 and 0.901, respectively. This work presents an interpretable and adaptive framework that provides competitive performance for churn prediction with a reduced feature set to provide valuable insights for improving customer retention strategies in the telecom industry.