<![CDATA[ABSTRAK Perilaku belajar mahasiswa berperan penting dalam menentukan performa akademik dan keberhasilan studi. Namun, identifikasi perilaku belajar masih banyak dilakukan secara manual sehingga kurang efektif untuk mendukung deteksi dini pada populasi mahasiswa yang besar. Penelitian ini bertujuan membangun model klasifikasi perilaku belajar mahasiswa menggunakan Random Forest Classifier (RFC) berbasis Learning Behavior Questionnaire (LBQ) dengan pendekatan CRISP-DM. Data dikumpulkan dari 400 mahasiswa menggunakan instrumen LBQ yang terdiri atas 28 item dalam tiga dimensi, yaitu Learning Behavior in Classroom, Learning Motivation, dan Social Interaction Outside Classroom. Instrumen memiliki validitas dan reliabilitas yang baik dengan nilai korelasi item 0,620–0,731 dan Cronbach’s Alpha 0,88–0,95. Untuk meningkatkan performa model, penelitian menerapkan SMOTE untuk mengatasi ketidakseimbangan kelas, Optuna untuk optimasi hiperparameter, dan SHAP untuk interpretasi model. Hasil pengujian menunjukkan bahwa RFC mencapai akurasi 98,33%, F1-Macro 0,9861, dan ROC-AUC 1,0000. Analisis SHAP mengidentifikasi IPK, motivasi intrinsik belajar (B8, B10, B9), serta perilaku belajar di kelas (A2, A5) sebagai prediktor paling berpengaruh. Temuan ini menunjukkan bahwa integrasi CRISP-DM, SMOTE, Optuna, dan SHAP menghasilkan model yang akurat, interpretatif, dan bebas data leakage, sehingga berpotensi digunakan sebagai sistem deteksi dini perilaku belajar mahasiswa di perguruan tinggi. ABSTRACT Student learning behavior plays an important role in academic performance and study success. However, learning behavior identification is often conducted manually, limiting its effectiveness for early detection in large student populations. This study aims to develop a student learning behavior classification model using a Random Forest Classifier (RFC) based on a Learning Behavior Questionnaire (LBQ) within the CRISP-DM framework. Data were collected from 400 undergraduate students using a 28-item LBQ covering three dimensions: Learning Behavior in Classroom, Learning Motivation, and Social Interaction Outside Classroom. The instrument demonstrated satisfactory validity and reliability, with item correlation coefficients ranging from 0.620 to 0.731 and Cronbach’s Alpha values between 0.88 and 0.95. The study employed SMOTE to address class imbalance, Optuna for hyperparameter optimization, and SHAP for model interpretability. The RFC model achieved 98.33% accuracy, 0.9861 macro F1-score, and 1.0000 ROC-AUC on the test set. SHAP analysis identified GPA, intrinsic learning motivation items (B8, B10, B9), and classroom behavior items (A2, A5) as the most influential predictors. These findings demonstrate that integrating CRISP-DM, SMOTE, Optuna, and SHAP produces an accurate, interpretable, and leakage-free model that can support early detection of student learning behavior in higher education.]]>
