The team worked with data from a computer fundamentals course delivered across four consecutive semesters at a Chinese university, a study formally approved by Swan College of Central South University of Forestry and Technology. The first three semesters supplied 24,440 student records as the source domain, each containing 36 attributes including scores from 26 practical training sessions. The fourth semester, taught by a different instructor with a substantially revised assessment scheme, formed the target domain with only 337 records, each carrying 79 attributes covering 42 practical sessions, four regular assignments, a group project and 22 unit tests. The goal was deliberately constrained and realistic: estimate a student’s standardized comprehensive score at the end of the course using only information available from the first eight practical training sessions, avoiding any label leakage from later assessments.

The first pillar of the framework is a semantic feature engineering step that translates wildly different assessment architectures into a common vocabulary. Rather than feeding raw session scores into the models, the researchers computed 13 temporally grounded statistical features from those first eight sessions, identically in both domains. These include the early mean, the standard deviation of early scores, a learning trend computed as the difference between the average of sessions five through eight and sessions one through four, counts of zero and non-perfect scores, early maximum and minimum, score range, pass rate, excellence rate, completion rate, the coefficient of variation, and a stability index defined as an exponential function of relative score dispersion. Because these definitions depend only on the first eight sessions, which exist in every semester, the resulting 13-dimensional feature spaces remain semantically comparable even though source courses had 26 sessions and the target course had 42 plus extra components.

Even with shared semantics, subtle distribution shifts persist across semesters because of differences in grading standards, assessment difficulty and student populations. The framework therefore applies a two-stage adaptation pipeline chosen specifically for the statistics of educational behavioral data. The authors observed, through a feature correlation heatmap, that the engineered features are strongly interdependent: standard deviation, score range and coefficient of variation all measure dispersion, while early mean, excellence rate and pass rate all reflect overall level. Applying per-feature corrections independently would distort these relationships. Instead, CORAL, short for CORrelation ALignment, first matches the covariance matrices of the source and target distributions, preserving the global correlation structure. A second stage then applies Maximum Mean Discrepancy, or MMD, guided per-feature weighting to fine-tune the residual mismatch, upweighting features whose post-CORAL discrepancies are larger.

The second pillar is a deliberately heterogeneous ensemble of 30 base models spanning five algorithmic families: eight tree-based variants drawn from Random Forest, Gradient Boosting and Extra Trees, six kernel-based Support Vector Regression variants, six linear models including Ridge, Lasso and Elastic Net, five multi-layer perceptron regressors with hidden architectures ranging from 32 to 128 neurons, and five K-Nearest Neighbor regressors with varied neighbor counts and distance metrics. This breadth of inductive bias means the ensemble captures complementary patterns, from nonlinear feature interactions to local similarity structure. The third pillar, and according to the ablation study the most critical one, is how these 30 predictions are fused: a Ridge-regularized stacking meta-learner treats the base predictions as meta-features and learns a weighted linear combination by minimizing regularized squared error, with the L2 penalty preventing the meta-learner from overfitting the few hundred target-domain samples.

Evaluation was designed to be rigorous. The team used 5-fold cross-validation with an 80/20 train/test split per fold on the 337 target records, yielding roughly 270 training and 67 test samples per fold, and repeated the procedure under 10 random seeds for a total of 50 fold-level trials. Hyperparameters were fixed in advance and never tuned on the target domain, and all distribution alignment was computed exclusively on training folds. The results were striking: the proposed framework achieved an R-squared of 0.5425 and a mean absolute error of 0.0851 on the target-domain test set, the best across all tested models, beating the strongest baseline, TabNet, by 4.09 percentage points in R-squared, an improvement the authors report as statistically significant under a paired t-test across the 50 trials.

The baseline comparisons carried their own lessons. Among individual learners, TabNet led with an R-squared of 0.5016, followed by Support Vector Regression trained on the target domain and LightGBM, whose histogram-based splitting proved efficient for tabular data. More dramatically, the domain adversarial neural network, or DANN, and a naive simple-averaging ensemble of the 30 base models both produced negative R-squared values, meaning worse than simply predicting the mean. The authors attribute DANN’s collapse to its gradient reversal layer destabilizing training when source and target distributions differ substantially in dimensionality and scale, while the failure of uniform averaging confirms that poorly transferring base models inject noise unless an adaptive mechanism learns to suppress them.

An ablation study quantified each component’s contribution by disabling modules one at a time under the same 50-trial protocol. Removing Ridge-regularized stacking in favor of simple averaging caused catastrophic degradation, confirming it as the single most important element. Cutting the ensemble down to only the eight tree-based models produced the second-largest drop, validating the value of cross-family diversity. Replacing the 13 engineered features with raw early-session scores moderately reduced accuracy and destabilized Spearman correlation. Interestingly, removing either CORAL or the MMD weighting did not degrade R-squared, which the authors attribute to semantically aligned label definitions, already-comparable features, and implicit compensation by the stacking layer, though the adaptation pipeline still reduced the overall cross-domain MMD by 26.92 percent, from 0.1230 to 0.0882, improving theoretical transferability for larger domain gaps.

Feature importance analysis added an intriguing practical finding. Using a random forest regressor to rank the engineered features, the team found a highly skewed distribution in which the excellence rate, the proportion of the first eight sessions scored at 90 or above, accounted for 73.88 percent of the importance among the top five features. Early mean, learning trend, coefficient of variation and stability index contributed the remainder. In plain terms, the strongest signal of final academic performance is not a single brilliant session but consistent early excellence, a conclusion that instructors designing early-warning systems can act on even without the full model.

The authors are candid about limitations: the target domain is a single semester of 337 students taught by one instructor, raising generalizability questions, although the 24,440-record source domain, the substantial structural differences between semesters, and the 50-trial evaluation protocol strengthen confidence. The framework’s promise is clear nonetheless. By combining a shared behavioral vocabulary, covariance-aware transfer and regularized ensemble fusion, it offers a practical template for educational early-warning systems in exactly the data-scarce settings where they are most needed, letting instructors identify at-risk students after just eight sessions, and future work extending validation to other courses, institutions and educational levels could determine how broadly this approach transforms learning analytics.

Subject of Research: Machine learning prediction of student academic performance across semesters using transfer learning and stacked ensemble methods

Article Title: A semantic-aligned stacked ensemble framework for cross-semester student performance prediction

Article References: Xiao, J., Han, P., Zeng, R., & Tan, Q. (2026). A semantic-aligned stacked ensemble framework for cross-semester student performance prediction. Discover Artificial Intelligence, 6(1), Article 1228. https://doi.org/10.1007/s44163-026-02328-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02328-3

Keywords: educational data mining, transfer learning, ensemble learning, student performance prediction, CORAL, maximum mean discrepancy, stacked generalization, domain adaptation, small-sample learning, Ridge regression, early-warning systems, machine learning