An Experimental Study on E-commerce Customer Churn Prediction Based on the Explainable XGBoost Algorithm

Abstract

The e-commerce industry has entered a phase of deepening market penetration, with high customer churn rates becoming a key issue constraining business development. Addressing the shortcomings of traditional churn prediction models—such as insufficient accuracy, poor interpretability, and failure to account for the heterogeneity of customer group needs—this paper proposes a framework for e-commerce customer churn prediction that integrates RFM value segmentation, the XGBoost algorithm, and SHAP interpretability analysis. Using a publicly available e-commerce dataset as the experimental sample, the RFM model was employed to segment customers into four value groups. Differentiated prediction models were constructed for each group, and SHAP was utilised to perform three-level interpretability analysis: global, group-level and local. Experimental results indicate significant heterogeneity in the churn drivers across the high-value, medium-value, potential-value and low-value customer sub-models. This study effectively addresses the ‘black box’ problem of models, offering enhanced interpretability and greater operational value, thereby providing support for e-commerce enterprises in making refined customer retention and operational decisions.