Multiclass Grade Prediction: Integrating Students’ Academic, Demographic, Self-Regulatory Learnings and Psychological Insights

Utilizing advanced machine learning approaches, we aim to predict students’ academic performance in an imbalanced multiclass classification setting. A dataset of 1,615 student records from Bangladeshi universities, with CGPA categorized into high, medium, and low groups, was used. The significant class imbalance posed a unique challenge (High:1491). To enhance prediction accuracy, we applied oversampling techniques (Random Oversampling, SMOTE) and two feature selection methods (Principal Component Analysis, Random Forest Gini importance). Eight machine learning models—Support Vector Classifier, Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest, AdaBoost, Extreme Gradient Boost, and Multiple Layer Perceptron—were trained using Stratified k-fold cross-validation and optimized with Bayesian hyperparameter tuning. The results demonstrated that Random Forest (f1 score: 0.937, roc-auc: 0.907) when combined with Random Oversampling and PCA feature selection, delivered the best performance, as assessed by the f1 score metric.