An Interpretable Machine Learning Framework for E-Commerce Customer Churn Prediction Using SHAP

Abstract

In the digital economy, the cost of customer acquisition for e-commerce platforms has been rising steadily. Customer retention has become an essential force for achieving sustainable profitability. Although the machine learning models exhibit superior performance in predicting customer churn, the black-box nature of some models limits their interpretability in practical applications, which in turn hampers their performance in real-world scenarios. This paper proposes a hybrid framework that combines advanced ensemble learning algorithms with interpretable artificial intelligence (XAI) techniques to bridge the gap caused by this limitation in machine learning. Multiple classification models were evaluated in this paper, including Logistic Regression, Random Forest, and XGBoost, in a high-fidelity synthetic e-commerce dataset. In order to bridge the gap between the model and business utility, the Shapley Additive exPlanations (SHAP) method is adopted to interpret the output of the optimal model. The results show that the proposed system can not only achieve stable predictive performance but also reveal the key drivers of customer churn, such as "engagement score" and "days since last purchase". Hence, the system can provide actionable, data-driven business strategies for customer retention.