<![CDATA[Predicting student academic success is an important problem in higher education due to the high risk of delayed graduation or dropout. However, the prediction process often faces challenges, such as an imbalanced class distribution, which can reduce the classification model’s performance. This study aims to evaluate and compare the performance of Logistic Regression, Random Forest, and Gradient Boosting algorithms in classifying student academic status into Graduate, Enrolled, and Dropout categories. The research method begins with data balancing using the Synthetic Minority Oversampling Technique (SMOTE) applied to the training data, followed by a comparison of two dimensionality reduction techniques: feature selection with Variance Threshold (VT) and feature extraction with Principal Component Analysis (PCA). Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. The results show that Random Forest without dimensionality reduction achieves the best performance, with an accuracy of 77.40% on SMOTE-balanced data. The application of SMOTE has also been shown to improve the balance of predictive capability across classes, as evidenced by higher F1-scores than with the original data, which is biased toward the majority class. Conversely, the use of PCA degrades model performance by reducing the information available to differentiate the classes. Furthermore, the Enrolled class is the most difficult to predict due to its high similarity in characteristics to the Graduate and Dropout classes. Based on these results, the combination of Random Forest and SMOTE, without dimensionality reduction, is the most effective approach for predicting students’ academic success. This model has the potential to be implemented as a component of an Early Warning System to support the identification of at-risk students and the implementation of earlier academic interventions. ]]>
Copyrights © 2026
