<![CDATA[Customer churn is a major challenge in the insurance industry because it directly affects customer retention, business sustainability, and company profitability. Early identification of customers at risk of churn is therefore essential for developing effective retention strategies. This study proposes an interpretable customer churn prediction framework for Sinarmas Syariah Lhokseumawe Insurance Services by integrating TabNet deep learning with Shapley Additive Explanations (SHAP). The dataset consists of 2,000 customer records containing demographic information, insurance transactions, premium payments, claims history, and customer interaction data. Due to the imbalanced class distribution, the Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the training dataset to improve model learning while preventing data leakage. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, and ROC-AUC metrics. The experimental results demonstrate that the proposed approach achieved an accuracy of 98.77%, precision of 86.67%, recall of 92.86%, F1-score of 89.66%, and ROC-AUC of 0.995, indicating excellent classification performance. Furthermore, SHAP analysis revealed that communication, premi_2025, and reason_to_purchase were the most influential features affecting churn predictions. These findings highlight the importance of customer engagement, premium management, and purchasing motivations in customer retention. The proposed TabNet-SHAP framework provides both high predictive performance and model interpretability, making it a valuable decision-support tool for customer retention strategies in the insurance sector.]]>
