Analysis of Factors Associated with Motor Vehicle Tax Arrears Using Ensemble Learning and Explainable Artificial Intelligence

Abstract

This study examines vehicle characteristics contributing to motor vehicle tax arrears classification in the context of UPTD PEPENDA Simalungun. The dataset contains 68,872 vehicles with three payment statuses. Binary analysis includes 66,887 records labelled as overdue or paid; 1,985 unpaid records are excluded pending clarification of their operational definition. Six predictors comprise brand, model, vehicle category, manufacturing year, registration plate colour, and tax amount. Data are partitioned by tax expiry year: 2023 for training, 2024 for validation, and 2025 for testing. Logistic regression provides a baseline for Random Forest and Histogram Gradient Boosting. Random Forest is selected using validation balanced accuracy of 0.5912 at a threshold of 0.60. Test ROC-AUC is 0.5452, balanced accuracy is 0.5356, overdue recall is 64.70%, and paid specificity is 42.41%. Interventional Shapley analysis of a test sample identifies manufacturing year, tax amount, and vehicle model as the largest absolute contributors. Discrimination remains limited, and feature contributions do not establish causal effects. Missing payment and extraction timestamps restrict the findings to exploratory payment-status classification rather than demonstrated prospective prediction.