Optimasi Algoritma XGBoost Berbasis SMOTE untuk Mengidentifikasi Faktor Logistik Pemicu Pembatalan Pesanan E-Commerce

Abstract

Unilateral order cancellation by consumers is a major operational challenge in the Indonesian e-commerce industry, directly impacting supply chain inefficiency and inflating logistics costs. Computational modeling to predict this risk often faces hurdles such as class imbalance and vulnerability to data leakage. This study proposes an optimization of the Extreme Gradient Boosting (XGBoost) algorithm integrated with the Synthetic Minority Over-sampling Technique (SMOTE) using a multi-scenario feature analysis approach. The dataset used comprises real transaction records from a major e-commerce platform. Experiments were designed in two scenarios: Scenario 1 included all transactional features, while Scenario 2 excluded the dominant financial feature (Total Pembayaran) to test the model's pure dependency on pre-finalization variables. The test results showed that Scenario 1 yielded a pseudo-accuracy of 99.50% due to data leakage. After reconstruction in Scenario 2, the SMOTE-based XGBoost model produced a stable real performance with an Accuracy Score of 83.96% and an Area Under ROC (AUC) of 86,98%. Through Feature Importance analysis, this study successfully revealed that pure logistics factors, specifically "Ongkos Kirim Dibayar oleh Pembeli" (Shipping Fee Paid by Buyer) with an absolute importance weight of 77.10%, serve as the primary predictor and driver behind consumer order cancellations. These findings provide tactical contributions for e-commerce decision-makers in formulating shipping cost strategies and mitigating early operational risks.