Customer Churn Prediction in Waste Banks Using XGBoost and SMOTE

Customer retention has become an important challenge in waste bank programs because declining member participation may reduce operational sustainability and weaken community-based waste management initiatives. However, churn prediction studies in non-commercial environmental programs such as waste banks remain limited. This study proposes a machine learning approach for customer churn prediction using operational transaction data from a waste bank managed by the Environmental Agency of Padang City, Indonesia. The dataset consisted of 34,188 transaction records representing 1,015 members collected between May 2024 and April 2026. Customer behavioral features were constructed from transaction history indicators, while class imbalance was handled using SMOTE and churn classification was performed using XGBoost under a leakage-aware customer-level train-test separation, ensuring a realistic evaluation on previously unseen members. Experimental results showed that the proposed model achieved an accuracy of 0.84, an F1-score of 0.73, and a ROC AUC value of 0.898. Feature analysis revealed that recency and transaction frequency were among the strongest predictors of churn behavior. The findings demonstrate the potential of machine learning to support participation monitoring and sustainability management in community-based waste bank systems.