Predicting customer repurchase intention through Electronic Word of Mouth analytics: a neural network approach

This study develops and theoretically grounds a Hybrid Convolutional–Dense Neural Network (HCDNN) to predict repurchase intention from electronic word-of-mouth (E-WOM) and structured behavioural data in the telecom sector. While prior research has extensively applied machine learning to churn prediction and sentiment analysis, limited attention has been devoted to theoretically interpreting predictive outputs within established consumer behaviour frameworks. Addressing this gap, this research integrates the Technology Acceptance Model (TAM) and Expectation–Confirmation Model (ECM) to explain how semantic signals embedded in E-WOM interact with behavioural attributes to shape repurchases intention. Using 5,200 annotated telecom reviews, the proposed HCDNN multimodal architecture achieves 94.5% accuracy and 97.0% ROC–AUC, outperforming classical machine learning algorithms and a transformer-based BERT baseline models. Ablation and cross-validation analyses confirm the robustness of the hybrid structure. The findings demonstrate that integrating semantic representation learning with dense behavioural modeling enhances both predictive performance and managerial interpretability. The study contributes by (1) bridging deep learning with behavioural theory, (2) advancing multimodal E-WOM analytics and (3) offering actionable insights for AI-enabled customer retention strategies. Future work will explore real-world CRM applications focusing on multilingual data, real-time deployment, and larger-scale implementation across the digital service industry.