An Ensemble Learning Framework for Student Academic Performance Prediction Using Hybrid Feature Selection and Hyperparameter Optimization

Abstract

The need for predicting students' academic performance is increasing nowadays since it is necessary for making decisions based on data and applying intervention for academically at-risk students. The problem is that the redundancy of features and poor configuration of the models reduce the accuracy of predictions in traditional machine learning algorithms. Therefore, the purpose of the present research was to design a hybrid ensemble algorithm for predicting students' academic performance based on the use of an e-learning dataset where Ant Colony Optimization (ACO) and Genetic Algorithm (GA) will be used for feature selection, Random Search for configuring classifiers' hyperparameters and XGBoost, AdaBoost, and Random Forest for classification. The proposed classifiers will be integrated through Majority Voting. The experiment has shown that the application of ACO-GA algorithm allowed identifying the best features while Random Search improved the performance of classifiers as it provided their proper configurations. While among individual models XGBoost showed the best classification results, the proposed ensemble classification model showed better results than any individual classifier and gave overall accuracy of 96.70%, Precision of 0.965, Recall of 0.963, F1-score of 0.964, and an ROC-AUC of 0.988. Thus, the comparison has revealed that the proposed framework provides higher predictive ability and lower classification errors than individual machine learning algorithms.