<![CDATA[Abstrak - Industri e-commerce di Indonesia menunjukkan perkembangan yang pesat dengan jumlah usaha mencapai 4,4 juta unit pada tahun 2024. Persoalan yang menyertai pertumbuhan tersebut adalah tingginya proporsi pesanan yang dibatalkan sehingga merugikan operasional penjual. Penelitian ini bertujuan membangun model klasifikasi guna memprediksi status akhir pesanan, yaitu Selesai atau Batal, melalui perbandingan tiga algoritma yaitu XGBoost, Random Forest, dan Decision Tree dengan kerangka CRISP-DM. Dataset bersumber dari Kaggle berjudul Indonesia E-Commerce Sales and Shipping 2023-2025 yang setelah pra-pemrosesan menyisakan 20.598 baris data. Permasalahan imbalance class dengan rasio 86,26% berbanding 13,74% diatasi melalui penerapan SMOTE pada data training. Tiga skenario dirancang yaitu tanpa SMOTE, dengan SMOTE, dan SMOTE yang dikombinasikan dengan hyperparameter tuning. Hasil eksperimen menunjukkan XGBoost tanpa SMOTE memperoleh kinerja terbaik dengan accuracy 89,73%, F1-Score 94,36%, dan AUC-ROC 76,99%. Decision Tree dengan SMOTE memberikan recall kelas Batal tertinggi sebesar 44,88%. Tiga atribut paling berpengaruh adalah jumlah barang, opsi pengiriman, dan metode pembayaran. Kata Kunci: Klasifikasi; E-Commerce; XGBoost; Random Forest; SMOTE; CRISP-DM; Abstract - The e-commerce industry in Indonesia shows rapid development with the number of businesses reaching 4.4 million units in 2024. The issue accompanying this growth is the high proportion of cancelled orders adversely affecting sellers' operations. This research aims to develop a classification model to predict the final status of orders, namely Completed or Cancelled, by comparing three algorithms: XGBoost, Random Forest, and Decision Tree using the CRISP-DM framework. The dataset originates from Kaggle titled Indonesia E-Commerce Sales and Shipping 2023-2025, which after preprocessing left 20,598 rows. The class imbalance problem with a ratio of 86.26% to 13.74% was addressed through SMOTE on training data. Three scenarios were designed: without SMOTE, with SMOTE, and SMOTE combined with hyperparameter tuning. Experimental results show XGBoost without SMOTE achieved the best performance with accuracy 89.73%, F1-Score 94.36%, and AUC-ROC 76.99%. Decision Tree with SMOTE provided the highest Cancelled class recall at 44.88%. The three most influential attributes are item quantity, shipping option, and payment method. Keywords: Classification; E-Commerce; XGBoost; Random Forest; SMOTE; CRISP-DM;]]>
