Comparative Evaluation of Hybrid Neural Network Architectures for Churn Prediction in Telecom Services

In the competitive telecom industry, customer retention is critical to sustaining profitability. Accurate churn prediction enables proactive customer retention strategies, helping telecom providers identify and address the risk of losing valuable customers. This research investigates the efficacy of four advanced hybrid neural network architectures—Simple Neural Network (SNN) + GRU + VRNN, LSTM + Random Forest, CNN + GRU, and Autoencoder + XGBoost—for predicting customer churn. By leveraging the complementary strengths of neural networks and traditional classifiers, these hybrid models aim to overcome the limitations of standalone architectures in handling complex, high-dimensional, and imbalanced telecom datasets. The study utilizes a comprehensive telecom dataset subjected to rigorous preprocessing techniques, including feature scaling, one-hot encoding, and Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. Each hybrid model is evaluated using key performance metrics such as accuracy, precision, recall, F1-score, and AUC-ROC. Results highlight the superior performance of hybrid models, with the Autoencoder +XGBoost combination achieving the highest accuracy of 88.1% and an AUC-ROC of 0.93. These findings underscore the potential of hybrid architectures in enhancing churn prediction, enabling telecom companies to deploy more effective retention strategies. This research contributes to the growing field of hybrid deep learning by providing a comparative analysis of models tailored to the telecom industry. The insights gained are expected to guide practitioners in selecting appropriate architectures for churn prediction and inspire future work in real-time implementations and further optimizations of hybrid systems.