A novel hybrid deep learning framework for customer churn prediction using RFM and embedding clustering

Abstract Customer churn prediction in E-commerce remains a challenging task due to the lack of labeled data, and the limited ability of traditional machine learning models to capture complex and dynamic customer behavior patterns. Existing approaches either rely on handcrafted features without effective representation learning or apply deep learning models without incorporating meaningful customer segmentation. To address these limitations, this study proposes a unified hybrid framework that integrates RFM-based feature engineering, Deep Embedded Clustering (DEC), and deep learning models for joint customer segmentation and churn prediction. The proposed framework first learns compact latent representations using a deep autoencoder, followed by clustering customers into behaviorally meaningful groups using an improved DEC mechanism with validated cluster selection and refinement. These learned representations are then used as input to Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) models for multi-class churn prediction. The framework is evaluated on two real-world datasets, namely the Online Retail dataset and an Events dataset, which differ significantly in scale and behavioral complexity. Experimental results demonstrate that LSTM achieves an accuracy of 99.65% on the Online Retail dataset and 99.83% on the Events dataset, while GRU achieves 99.77% and 99.75%, respectively. Although traditional models such as Logistic Regression and Support Vector Machine achieve competitive performance, they show limited adaptability across heterogeneous data distributions. The results confirm that integrating representation learning with clustering and deep sequential models significantly enhances churn prediction performance. The proposed framework effectively transforms raw transactional data into structured, actionable insights that support customer retention strategies in E-commerce environments.