Optimizing Multi-Class Classification in Educational Data with Ensemble Learning and Data Balancing Techniques

The field of education is increasingly embracing AI tools to improve student outcomes. This work aims to reduce academic failure in higher education by employing machine learning techniques to identify at-risk students early in their educational journey, enabling the implementation of supportive strategies to assist them. This study examines a dataset from a higher education institution and utilizes it to develop a classification model for predicting students’ academic performance. The problem is formulated as a multi-class classification task with three categories: Graduate, Enrolled, and Dropout, with a significant imbalance skewed toward the Graduate. To improve prediction accuracy toward the minority class, the data balancing technique SMOTE with Edited Nearest Neighbor (SMOTE-ENN) is applied. Three popular classification models—Random Forest, XGBOOST and CatBoost—are employed. The findings show that SMOTE-ENN significantly improves classification results. Moreover, XGBOOST demonstrated the highest accuracy (94.6%) in correctly identifying all classes, as evidenced by the confusion matrix evaluation, achieving the highest results compared to previous work in the literature. Implementing these models allows for accurate predictions of students’ performance and helps reduce dropout rates.