<![CDATA[Student academic performance prediction is an essential component of Educational Data Mining (EDM) for the early identification of at-risk students. This approach aims to improve prediction accuracy and support decision-making for the early identification of at-risk students. This study proposes an optimized prediction pipeline that integrates feature engineering, Pearson Correlation-based Feature Filtering (PCFF), the Synthetic Minority Over-sampling Technique (SMOTE), and Random Forest (RF) to predict student academic performance. The Portuguese Student Performance Dataset (1,043 clean records; Pass = 814, Fail = 230) was used for evaluation. Four engineered features were constructed, reducing the feature space to 15 features through PCFF (threshold |r| ? 0.1). SMOTE was applied exclusively within each training fold to prevent data leakage. Two primary models were evaluated: a baseline Naïve Bayes model (Accuracy = 90.43%) and the proposed RF Default + SMOTE model (Accuracy = 93.78%, Recall = 95.09%, F1-score = 95.98%). Ten-fold stratified cross-validation achieved an accuracy of 90.89% ± 2.08%. The engineered feature G_avg obtained the highest feature importance score (0.285), outperforming the original grade features. The results demonstrate that integrating SMOTE and feature engineering significantly improves minority class detection, reducing False Negatives from 15 (baseline) to 8 (RF + SMOTE), representing a 46.7% improvement in identifying at-risk students.]]>
Copyrights © 2026
