A hybrid feature selection framework for dimensionality-optimized customer churn prediction.

Customer churn is a major problem in the fiercely competitive telecom industry. It directly influences the profitability of a company and its future in the industry. There have already been successful applications of machine learning and deep learning (ML/DL) techniques to detect client churn. learning (ML/DL) techniques to detect client churn. However, in certain situations, ML/DL-based algorithms fall short of providing encouraging outcomes when it comes to classifying customer churn. The use of machine learning classifiers and conventional feature encoding techniques in earlier studies on customer turnover estimation produced promising predictions. But the role of the optimized feature set remains important for improving efficiency and reducing feature dimension. This research proposes a modular churn prediction framework that integrates data preparation, feature optimization through hybrid feature selection, class balancing, and predictive modeling into a unified pipeline. This is done through a Multi-stage Hybrid Feature Selection (MHFS) technique that combines Mutual Information (MI) for preliminary relevance filtering, Recursive Feature Elimination (RFE) for iterative refinement, Boruta for rigorous variable validation, and Parallel Simulated Annealing (SA) for global optimization. Furthermore, a 5 × 3 nested stratified cross-validation is used to guarantee objective model evaluation where feature selection and SMOTE are performed exclusively within inner training folds, while outer folds remain completely unseen for performance assessment. Multiple ML algorithms are used for training, including Random Forest, XGBoost, LightGBM, SVM, Extra Trees, and Logistic Regression, along with Tabular Neural Network. The framework is evaluated using two comprehensive telecommunications datasets, Cell2Cell and IBM Telco. It will ensure the efficacy of the proposed framework across different churn settings datasets, which is then combined with nested cross-validation on the training set. The experimental results show that the proposed hybrid framework achieves competitive predictive performance while substantially reducing feature dimensionality (from 28 to 16 features in IBM Telco and from 57 to 25 features in Cell2Cell). Using XGBoost, the framework achieves an accuracy of 81.60% on IBM telco and 77.52% on the cell2cell dataset. These results show that the proposed framework provides competitive performance while significantly reducing feature dimensions.