Abstract:

Customer churn threatens recurring revenue in subscription services, motivating reliable identification of subscribers at risk of leaving. This study evaluates logistic regression, Random Forest, XGBoost and a multilayer perceptron (MLP) on 992,931 KKBox subscribers. A scalable preprocessing workflow integrates demographic records, transactions and listening logs into 35 predictors, including recency, activity within successive 30-day windows and 90-day aggregates. Models were assessed on a common stratified hold-out set (n = 198,587, approximately 6.4% churn), with imbalance-aware learning. XGBoost yielded the best overall balance of performance (accuracy 0.96, ROC-AUC 0.98, precision 0.65 and F1-score 0.73). Random Forest obtained the highest recall (0.88). Classification accuracy was close to that reported for a substantially more complex hybrid graph attention–MLP model on KKBox data. The tabular XGBoost approach avoids graph construction and attention-based message passing, suggesting a simpler prediction architecture, although runtime savings were not experimentally measured. Transaction recency and cancellation history were consistently informative. The findings favour carefully engineered behavioural features and efficient nonlinear classification, subject to limitations of single-dataset evaluation and non-identical published benchmarks.