Leveraging RFM Features for Churn Prediction in Imbalanced E-Commerce Data: A Hepsiburada Case Study

The share of the e-commerce sector in the global economy and the level of market competition are rapidly increasing. Retaining existing customers generally incurs lower costs than acquiring new ones, making churn prediction a strategic priority for e-commerce platforms. This study introduces a framework that leverages RFM (Recency, Frequency, Monetary) scores calculated across 32 product categories to construct symbolic behavioral feature vectors for churn prediction. Unlike traditional RFM segmentation, these categorical scores are directly used as inputs for supervised machine learning models, enhancing interpretability and enabling fine-grained behavioral analysis. A binary churn label was generated based on purchase activity over the final two months of a 10-month period in 2024, using a real-world e-commerce dataset from HepsiBurada.com, revealing significant class imbalance (82% active, 18% churned). To address this, seven resampling strategies - including SMOTE variants and hybrid techniques - were applied in combination with five classifiers: Decision Tree, Random Forest, Support Vector Machine, XGBoost, and Artificial Neural Network. A total of 35 model-resampling combinations were evaluated using accuracy, F1-score, recall, and confusion matrices. Random Forest combined with Random Oversampling achieved the highest performance, with 95% accuracy and 94% F1-score. The representation of RFM-based features proved particularly robust under imbalanced conditions of a real-world e-commerce dataset from HepsiBurada.com. The proposed framework demonstrates the effectiveness of integrating RFM-based segmentation with predictive modeling for real-world e-commerce scenarios and provides actionable insights for businesses aiming to proactively identify at-risk customers and reduce churn.