Mitigating domain shift in cross-domain student performance prediction using a CORAL-based adaptation framework

Student performance prediction has become an important research area in educational data mining because it supports early academic interventions and decision-making processes. However, most previous studies assume that training and testing data originate from the same academic environment, which limits model generalization when applied to different academic programs. In practical educational settings, variations in student characteristics, academic behavior, and learning distributions across programs introduce substantial cross-domain distribution shifts that can significantly reduce prediction performance. This study proposes a CORAL-based cross-domain learning framework for student performance prediction under shifting distribution conditions. The framework applies CORrelation ALignment (CORAL) domain adaptation to align the feature covariance between the source and target programs before model training and evaluation. Four machine learning models, namely logistic regression (LR), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGB), were evaluated using three academic program datasets under within-domain, conventional cross-domain, and CORAL-based adaptation settings. Performance was assessed using accuracy, precision, recall, F1-score, and area under the curve (AUC), with the F1-score used as the primary evaluation metric. The experimental results demonstrate that baseline cross-domain learning consistently experiences performance degradation owing to distribution shifts. The proposed CORAL-based framework successfully improved the cross-domain prediction performance in most experimental scenarios and significantly reduced the performance gap toward within-domain learning. The highest improvement was achieved by the random forest model in the ASY target domain, where the F1-score increased from 0.737 to 0.802 after adaptation. Statistical analysis using the Wilcoxon signed-rank test further confirmed that most of the improvements were statistically significant. The findings indicate that CORAL-based domain adaptation provides an effective and computationally efficient solution for improving cross-domain student performance prediction under heterogeneous educational data distribution.