Wiyli Yustanti, Ricky Eka Putra, I. Gusti Putu Asto Buditjahjanto, Andi Iwan Nurhidayat
Timely graduation is a key performance indicator in higher education and a critical goal for students, institutions, and policymakers. This study investigates the effectiveness of various machine learning algorithms in predicting students' graduation timeliness using multidimensional data, which includes both academic and non-academic attributes. A comparative analysis was conducted using classification models, including Random Forest, XGBoost, LightGBM, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naive Bayes, Logistic Regression, and ensemble methods such as Bagging, Voting, and Stacking. To handle class imbalance and enhance model robustness, class weighting, SMOTE (Synthetic Minority Oversampling Technique), and hyperparameter tuning were applied. The experimental results show that LightGBM consistently outperformed all other models, achieving the highest test accuracy (0.72), ROC AUC (0.71), and macro-average F1-score (0.64). This indicates LightGBM's ability to effectively balance precision and recall for both the ontime graduation and not on-time graduation classes. Other ensemble methods, such as Random Forest and Stacked Models, also demonstrated competitive results. In contrast, traditional models like Naive Bayes and Logistic Regression underperformed, particularly in identifying students at risk of delayed graduation. These findings confirm the potential of LightGBM as a reliable model for predicting graduation outcomes and highlight the importance of integrating both academic and socioeconomic factors in educational data analytics. © 2025 IEEE.
Universitas Negeri Surabaya, Master's Study Program in Informatics, Surabaya, Indonesia
