Customer Churn Prediction Uses Machine Learning to Improve Retention on Digital Platforms

Abstract

Customer churn is a critical challenge for digital platforms operating in highly competitive markets such as e-commerce. This study aims to develop a machine learning–based predictive model to identify Shopee customers in Indonesia who are at high risk of churn, using behavioral and transactional data. A supervised learning approach was employed using multiple algorithms, including Logistic Regression, Decision Trees, Random Forests, and XGBoost. The dataset consisted of user activities, including transaction frequency, recency, voucher usage, application session count, and interaction with promotional features. Data imbalance was addressed using the SMOTE technique to improve classification stability. Results showed that XGBoost achieved the best performance across all evaluation metrics, with an AUC of 0.948, indicating strong discriminative ability. Feature importance analysis revealed that recency, transaction frequency, voucher usage rate, and app session frequency were the most influential predictors of churn. These variables indicate declining engagement and reduced responsiveness to promotional incentives, which are key behavioral signals of churn. The study contributes to both academic literature and practical applications by demonstrating how behavioral analytics and machine learning can support early churn detection and inform targeted retention strategies. Implementing such predictive systems can help e-commerce platforms optimize customer lifetime value and reduce revenue loss.