Perbandingan Algoritma Supervised Learning dalam Memprediksi Jalur Seleksi Masuk Mahasiswa di Universitas Negeri Gorontalo

Abstract

The choice of university admission pathway (SNBT, SNBP, or independent selection/Mandiri) is closely related to students' demographic and administrative characteristics, and understanding this pattern benefits higher education institutions in designing more targeted recruitment strategies. This study aims to compare the performance of four supervised learning algorithms Decision Tree, Random Forest, Naïve Bayes, and k-Nearest Neighbor (k-NN)  in predicting the admission pathway of Informatics Engineering students at Universitas Negeri Gorontalo (UNG). Data were obtained from UNG's Integrated Academic Information System (SIAT) for 1,237 students from the Information Technology Education and Information Systems study programs (2018–2024 cohorts), which after data cleaning resulted in 1,232 samples across three pathway classes (SNBT = 682, SNBP = 417, Mandiri = 133). Feature selection using information gain identified five informative features: selection type (national/local), age, cohort year, initial registration semester, and gender. Evaluation was conducted through three scenarios: 80:20 hold-out with Synthetic Minority Oversampling Technique (SMOTE), and 10-fold stratified cross-validation, followed by a Friedman significance test. The 10-fold cross-validation results show mean accuracies of 64.61% for Random Forest, 64.37% for Decision Tree, 63.47% for Naïve Bayes, and 61.03% for k-NN. The Friedman test indicated no statistically significant difference in performance among the four algorithms (χ² = 4.39; p = 0.222). The confusion matrix revealed that all algorithms perfectly classified the Mandiri class due to a definitional dependency with the selection-type feature, while most misclassifications occurred between the SNBP and SNBT classes. Applying SMOTE consistently improved minority-class (SNBP) recall across all four algorithms, but its effect on macro F1-score and overall accuracy varied by algorithm and was not uniformly positive. This study recommends incorporating academic behavioral features to improve the model's discriminative ability between national selection pathways.