Deep Multi-level Ensemble Model for Customer Churn Prediction

Customer churn prediction is a critical challenge in customer analytics, especially in highly competitive sectors such as telecommunications and e-commerce. While ensemble learning has shown promise in improving classification accuracy, traditional stacking methods often fail to capture deep interactions between base learners. This study proposes a novel multi-level ensemble model, XMS-Net, which combines XGBoost, LightGBM, and multi-layer perceptron classifiers using a two-level stacking architecture with a deep MLP meta learner. The model is evaluated on four real-world datasets. Experimental results show that XMS-Net significantly outperforms baseline classifiers and traditional stacking methods across multiple metrics, including F1-score, accuracy, recall, precison, and ROC-AUC. Highest accuracy score achives 97, improvements of up to 6.8 compared to the strongest single model are reported. Statistical tests (paired t-test and Wilcoxon signed-rank test) confirm the significance of these gains. An ablation study further highlights the importance of integrating both gradient boosting and neural learners. These findings indicate that XMS-Net provides a robust and scalable approach for churn prediction and may be generalized to other domains involving high-dimensional, imbalanced data. The study contributes a validated ensemble framework for enhancing predictive performance in real-world classification tasks.