<![CDATA[Customer retention is a critical challenge in the e-commerce industry, yet many platforms frequently suffer from a lack of explicit labels to identify potential defectors (churn). This research proposes a hybrid unsupervised-supervised learning framework to perform automated churn classification using the high-dimensional Olist Brazilian E-commerce dataset. The first stage employs K-Means Clustering to objectively generate churn labels based on Recency, Frequency, and Monetary (RFM) features. The second stage applies an Extreme Gradient Boosting (XGBoost) model, augmented by the Synthetic Minority Over-sampling Technique (SMOTE), to address the inherent class imbalance typical of transactional data. Experimental results demonstrate that the proposed model achieved a robust accuracy of 72% with a churn recall rate of 69%. Interpretability analysis using SHapley Additive exPlanations (SHAP) revealed that delivery duration and customer review scores are the most dominant predictors of churn, significantly outweighing financial metrics. These findings contribute a novel integration of automated labeling and model transparency, enabling e-commerce managers to implement proactive, data-driven customer retention strategies.]]>
