Machine learning models that predict whether students will graduate on time often stumble when they are moved from one academic program to another, and a new study offers a lightweight statistical fix. Researchers report that a technique called CORrelation ALignment, or CORAL, can substantially reduce the performance loss that occurs when prediction models trained on one study program are applied to a different one. The work, published in Discover Artificial Intelligence, addresses a persistent weakness in educational data mining.
Most student performance prediction studies train and test their models on data drawn from the same academic environment, a setup known as within-domain learning. That assumption produces impressive benchmark numbers but breaks down in practice, because smaller or newly established study programs often lack the historical records needed to build reliable models of their own. When a model trained on one program is applied to another, differences in student characteristics, curricula, and academic behavior create a domain shift that degrades predictions.
The research team, led by Achmad Bisri and colleagues, treated three study programs, labeled PBS, ESY, and ASY, as separate domains. The combined dataset contained 276 student records, split into 104, 126, and 46 instances per program, with 97 students graduating on time and 179 experiencing delays. The ASY program, with only 46 records, served as a realistic low-resource target domain.
Each record included demographic attributes such as gender, school of origin, and admission pathway, alongside age and grade point averages from the first four semesters, with graduation status as the label to predict. To quantify the problem, the researchers evaluated four classifiers, logistic regression, support vector machine, random forest, and extreme gradient boosting, under three settings. Within-domain learning, using stratified cross-validation on a single program, represented the upper bound.
Conventional cross-domain learning trained models on the combined remaining programs and tested them directly on the target without adaptation. The third setting applied CORAL before training. Performance was measured with accuracy, precision, recall, F1-score, and area under the curve, with the F1-score, which balances precision and recall, serving as the primary metric.
The baseline experiments confirmed that domain shift is real and uneven. In the ASY domain, the support vector machine's predicted F1-score fell from 0.938 within-domain to 0.783 cross-domain, while random forest dropped from 0.926 to 0.737, the largest F1 decline of 0.189 among all models. The PBS domain showed smaller degradation, with random forest losing only 0.032.
Intriguingly, the ESY domain bucked the trend: every model scored slightly better cross-domain than within-domain, with extreme gradient boosting gaining 0.082. The authors suggest that the larger, more diverse combined source data provided richer decision boundaries than ESY's own limited records, indicating that the impact of domain shift depends on how compatible the underlying distributions are. CORAL works by aligning the second-order statistics of the source and target feature spaces.
The method computes covariance matrices for both domains, then transforms the source features in two steps: a whitening stage that removes the source covariance structure, followed by a recoloring stage that imposes the target's covariance structure. Crucially, CORAL requires no labeled data from the target program, preserves the original feature dimensionality, and plugs into standard machine learning pipelines without altering the classifiers themselves. That makes it far cheaper computationally than deep domain adaptation methods, which the authors note were designed for large-scale image and signal datasets rather than the small, structured, tabular records typical of education.
The alignment measurably worked. Centroid distances between combined source and target distributions fell from 0.71 to 0.23 for PBS, a 67.46 percent reduction; from 0.94 to 0.25 for ESY, a 72.99 percent reduction; and from 1.34 to 0.25 for ASY, an 80.99 percent reduction. Principal component analysis visualizations showed source and target samples overlapping far more after adaptation.
The largest initial discrepancy, in ASY, yielded the greatest correction, suggesting CORAL is most valuable precisely where the shift is most severe. Performance gains followed the same pattern. In ASY, random forest improved by 0.065 and extreme gradient boosting by 0.050, both highly significant with p-values below 0.001, while logistic regression gained 0.007.
In PBS, the support vector machine improved by 0.025 to reach the domain's best adapted F1-score of 0.830, and logistic regression gained 0.021. In ESY, where cross-domain performance was already strong, gains were marginal but still statistically significant across all models. A recovery rate analysis, which measures how much of the within-domain to cross-domain performance gap is closed, found the support vector machine recovered 0.469 of the gap in PBS and random forest recovered 0.344 in ASY.
Not every combination benefited: the support vector machine in ASY and random forest and extreme gradient boosting in PBS showed slight declines, because CORAL aligns only covariance and does not explicitly preserve class boundaries or higher-order distribution structure. The statistical validation reinforced these conclusions. Using the Wilcoxon signed-rank test on paired F1-scores from 30 repeated runs with 5-fold stratified cross-validation, the researchers found that CORAL significantly outperformed baseline cross-domain learning in most model-domain pairs, with the exception of the non-significant support vector machine result in ASY and extreme gradient boosting in PBS.
Comparisons against the within-domain upper bound showed that significant gaps remained in ASY and PBS, meaning adaptation narrowed but did not eliminate the cost of transfer. In ESY, however, CORAL-adapted cross-domain models significantly outperformed within-domain models for logistic regression, random forest, and extreme gradient boosting, demonstrating that knowledge transfer can sometimes exceed what a data-poor program achieves alone. The authors are careful about scope.
The ASY program's small sample may affect the stability of performance estimates, and the study relied on demographic and early-semester academic variables because behavioral data such as attendance or learning management system activity were not consistently available across programs. Because CORAL depends on covariance estimates, noisy or irrelevant features could weaken the alignment. Still, the framework's implications extend beyond graduation prediction: the same covariance alignment could apply to dropout prediction, retention analysis, or any educational task where institutions with thin data need to borrow statistical strength from related programs, provided comparable predictor variables exist across domains.
Subject of Research: CORAL-based domain adaptation for cross-domain student performance prediction across academic programs
Article Title: Mitigating domain shift in cross-domain student performance prediction using a CORAL-based adaptation framework
Article References: Bisri, A., Yuhana, U. L., Amaliah, B., & Anwar, M. (2026). Mitigating domain shift in cross-domain student performance prediction using a CORAL-based adaptation framework. Discover Artificial Intelligence, 6(1), Article 1651. https://doi.org/10.1007/s44163-026-02002-8
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02002-8
Keywords: student performance prediction, domain adaptation, CORAL, educational data mining, machine learning, domain shift, random forest, cross-domain learning, graduation prediction, covariance alignment, higher education, Wilcoxon signed-rank test
News Source: Denise Maddox. (October 26, 2026). Covariance Alignment Boosts Cross-Program Student Performance Prediction, Study Finds. Scienmag.
