<![CDATA[Abstrak - Penerapan machine learning pada data administratif memiliki potensi besar dalam mendukung analisis berbasis data, namun performa model sangat dipengaruhi oleh karakteristik dataset yang digunakan. Penelitian ini bertujuan untuk menganalisis kinerja algoritma Support Vector Machine (SVM), Random Forest, dan XGBoost dalam melakukan klasifikasi tingkat pendidikan warga binaan berdasarkan data administratif non-identitas. Dataset diperoleh dari arsip administrasi internal lembaga pemasyarakatan dan melalui tahapan pra-pemrosesan, pengubahan fitur kategorikal menggunakan metode one-hot encoding, serta pemisahan data latih dan data uji dengan rasio 80:20. Evaluasi performa model dilakukan menggunakan metrik akurasi, presisi, recall, dan F1-score. Hasil pengujian menunjukkan bahwa akurasi berada pada kisaran 0,15–0,28, presisi 0,04–0,09, recall 0,07–0,10, serta F1-score 0,06–0,10. Model SVM dengan kernel Radial Basis Function (RBF) memperoleh nilai akurasi tertinggi, sementara XGBoost menunjukkan kinerja yang relatif lebih stabil pada metrik presisi, recall, dan F1-score. Namun demikian, capaian performa tersebut hanya sedikit lebih baik dibandingkan baseline klasifikasi acak, sehingga secara keseluruhan kinerja model masih tergolong rendah. Temuan ini menunjukkan bahwa dominasi atribut kategorikal, keterbatasan informasi fitur, serta ketidakseimbangan distribusi kelas memberikan pengaruh signifikan terhadap kemampuan model dalam mempelajari pola secara optimal. Kata kunci: machine learning; evaluasi kinerj;, data administratif; klasifikasi; warga binaan; Abstract - The application of machine learning to administrative data has great potential in supporting data-driven analysis, but model performance is greatly influenced by the characteristics of the dataset used. This study aims to analyze the performance of Support Vector Machine (SVM), Random Forest, and XGBoost algorithms in classifying the educational level of inmates based on non-identifying administrative data. The dataset was obtained from the internal administrative archives of correctional institutions and underwent pre-processing, categorical feature conversion using the one-hot encoding method, and separation of training and test data with a ratio of 80:20. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. The test results showed that the accuracy ranged from 0.15 to 0.28, precision from 0.04 to 0.09, recall from 0.07 to 0.10, and F1-score from 0.06 to 0.10. The SVM model with Radial Basis Function (RBF) kernel obtained the highest accuracy value, while XGBoost showed relatively more stable performance in terms of precision, recall, and F1-score metrics. However, these performance achievements were only slightly better than the random classification baseline, so overall model performance was still relatively low. These findings indicate that the dominance of categorical attributes, limited feature information, and class distribution imbalance have a significant impact on the model's ability to learn patterns optimally. Keywords: machine learning; performance evaluation; administrative data; classification; inmates;]]>
