Comparative analysis of ensemble learning algorithms in enhanced confidence-based assessments

Abstract

This paper provides a comparative analysis of ensemble learning (EL)algorithms to enhance the confidence-based assessment (CBA) in evaluating student performance. Traditional CBA often suffers from misclassification caused by overconfidence and underconfidence, limiting its accuracy and fairness. To address these challenges, an enhanced CBA-EL model integrating bagging and boosting ensemble algorithms is proposed. Five bagging algorithms, which are random forest (RF), decision tree (DT)support vector machine (SVM), K-nearest neighbors (KNN), Naïve Bayes(NB), and four boosting algorithms, which are adaptive boosting(AdaBoost), eXtreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost), were evaluated using a dataset of 276 responses collected from Pre- and Post-Quiz CBA in a discrete structures course. Algorithm performance was evaluated using accuracy, correlation, weighted mean precision (WMP), and weighted mean recall (WMR). RF achieved 73.19% accuracy, 0.725 correlation, 0.751 WMP, and 0.766 WMR, while CatBoost outperformed all with 86.23% accuracy and the highest correlation, WMP, and WMR values, with 0.842, 0.843, and 0.862, respectively. The findings indicate that integrating EL into CBA improves prediction accuracy and supports bias-aware student evaluation. This research advances reliable assessment practices and informs the development of adaptive learning systems.