The study drew on the Open University Learning Analytics Dataset, or OULAD, a rich public resource containing demographics, registration records, module metadata, assessment scores, and interaction logs from a distance-learning institution. The researchers started with 32,593 enrolments and applied careful exclusion rules. Learners who had unregistered on or before the early prediction cutoff were removed, because their outcome was already determined by withdrawal before the prediction point. Enrolments also needed at least one valid non-banked continuous assessment score. The final analytical cohort comprised 24,153 enrolments spanning 22 module presentations across seven modules and 22,234 unique learners.

The prediction window was deliberately strict. Early engagement was restricted to virtual learning environment interactions occurring on or before 25 percent of each module presentation length, corresponding to roughly days 58 to 67 across presentations. The target was a continuous score on a 0 to 100 scale, calculated as the weighted mean of tutor-marked assessments and computer-marked assessments. Examination scores were excluded from target construction, banked scores carried over from earlier attempts were removed, and no assessment score contributing to the target was allowed as a predictor variable. These leakage controls matter enormously in learning analytics, where subtle contamination between features and outcomes can inflate apparent accuracy.

From this foundation, the team constructed 84 predictors, 78 numerical and 6 categorical, drawn from three domains. Learner-level predictors included demographic and registration attributes plus early-engagement measures such as total clicks, number of active days, distinct learning materials accessed, timing of first and last activity, and activity-type-specific click counts. Course-design predictors described the composition of virtual learning environment materials, the number and spacing of assessments, assessment weights, and presentation length. Course and presentation identifiers were deliberately excluded from the predictor matrix because they defined the grouped validation units, preventing the model from simply memorizing course-specific quirks.

The proposed DKBNE-AlexNet architecture is a genuinely unusual hybrid. It combines a Deep Kronecker Network branch, a feature-token batch-normalized eLU AlexNet-inspired branch, gated residual fusion, and a wide linear residual path. The Deep Kronecker Network offers parameter-efficient representation learning by expressing each weight matrix as a sum of Kronecker products, dramatically reducing the number of free parameters compared with an unrestricted dense matrix. In the final configuration, the DKN branch processes input through 256 and 128 dimensional Kronecker layers with eLU activation, LayerNorm, dropout, and residual projection to a 96-dimensional latent representation.

The second branch is an inventive adaptation of AlexNet, the classic convolutional architecture originally designed for two-dimensional images. Rather than reshaping the feature vector into an artificial image, which would impose arbitrary spatial adjacency on heterogeneous tabular columns, each scalar predictor was first converted into a feature-specific 24-dimensional token. Five 1 by 1 Conv1D stages with batch normalization and eLU activation then transform these feature tokens without mixing neighboring columns. Attention, mean, and max pooling are combined to form a global representation projected onto the fusion space. This design preserves a five-stage convolutional transformation while respecting the tabular nature of the data.

Fusion and training are equally thoughtful. The DKN and convolutional representations are projected to a common space and combined through their direct values, element-wise interaction, and absolute difference. A learned sigmoid gate scales the residual correction produced by the convolutional and fusion path, while a wide linear path from the processed original predictors is added so that simple additive relationships can be represented directly. The model trains in two stages: first the DKN anchor is optimized, then the convolutional and fusion branch is trained as a residual correction while the anchor is frozen. This allows the inner validation procedure to retain the anchor when an added correction is not beneficial. Training used the Huber loss with delta 0.75, the AdamW optimizer, a batch size of 512, an initial learning rate of 7 times 10 to the power of negative 4, and early stopping with patience of 20 epochs.

Generalization was evaluated with unusual rigor. Five-fold cross-validation grouped by module presentation was repeated three times, so all enrolments belonging to a held-out presentation were assigned exclusively to the test fold. This prevented the same course presentation from contributing records to both training and testing and evaluated performance on course designs not observed during model fitting. Fifteen outer evaluations were completed for every model. Performance was assessed using mean absolute error, root mean squared error, and the coefficient of determination, with MAE as the primary metric because it is directly interpretable in score points. Uncertainty was estimated using a group-aware bootstrap over module presentations with 5,000 resamples, and model comparisons used paired Wilcoxon signed rank tests with Holm adjustment for multiple comparisons.

The results tell a nuanced story. DKBNE-AlexNet achieved an MAE of 11.424, with a 95 percent group-bootstrap confidence interval of 10.389 to 12.560, an RMSE of 14.843, and an R squared of 0.117. Random Forest achieved the lowest pooled MAE at 10.916, followed closely by LightGBM at 10.918, while the BN-eLU-AlexNet component alone achieved 11.173 and the DKN component alone 11.416. Crucially, the presentation-level difference between DKBNE-AlexNet and Random Forest was not statistically significant, with a Wilcoxon p value of 0.388 and a Holm-adjusted p value of 1.000. The proposed model also displayed a compact repeated-fold MAE distribution with a fold-level standard deviation of approximately 0.61, indicating improved stability across held-out presentations.

Feature attribution revealed what actually drives predictions. Held-out group-aware permutation importance identified the number of distinct learning materials accessed as the largest contributor, with a mean importance of 0.329 MAE points and a Holm-adjusted p value below 0.001. Highest previous education followed at 0.208, then active learning days at 0.168, forum engagement at 0.103, and the IMD group, a measure of neighborhood deprivation, at 0.086. The timing of the first recorded interaction also remained significant after correction. Individual course-design variables generally showed smaller and less stable effects. The message is clear: the diversity and regularity of early engagement matter more than static course structure.

The authors are refreshingly candid about limitations. The target is a learner assessment outcome and cannot isolate teaching quality from learner preparedness, course difficulty, and other contextual influences. OULAD does not provide tutor identifiers, so instructor effects cannot be estimated separately from presentation effects. The analysis covers 22 module presentations from a single distance-learning institution, limiting external generalization. Residual analysis shows regression toward the mean at extreme scores, with very low observed scores overpredicted and very high observed scores underpredicted. The ablation study did not establish a clear average-error advantage from combining the two branches, and permutation importance identifies predictive associations rather than causal effects.

Perhaps the most important implication concerns fairness and deployment. Because highest education and IMD bands contribute predictive information, any operational deployment should include a fairness review and avoid using model output as a punitive or high-stakes decision about individual learners or instructors. The framework is best interpreted as an early-warning learning-analytics instrument rather than a direct evaluation of teacher effectiveness. Predictions constructed during the initial 25 percent of a module presentation can inform targeted instructional interventions, academic support prioritization, or systematic monitoring of presentations exhibiting elevated outcome risk.

For the learning analytics community, the study offers a competitive and stable neural framework while emphasizing that strong tree-based baselines remain essential. The overlapping confidence intervals between the hybrid model and Random Forest suggest that both branches contain useful predictive information, even if fusion did not improve average accuracy in this experiment. Future work could complement permutation-based attribution with instance-level explainable AI methods to improve the transparency of individual predictions. As online education continues to expand worldwide, tools like DKBNE-AlexNet point toward platforms that watch not just how much students click, but how broadly and consistently they explore, and that reach out with support while there is still a semester left to save.

Subject of Research: Early prediction of learner assessment outcomes in e-learning using a hybrid deep Kronecker and BN-eLU-AlexNet framework

Article Title: A hybrid deep Kronecker and BN-eLU-AlexNet framework for early prediction of learner outcomes in e-learning platforms

Article References: Malkani, C., Mittal, N., & Som, S. (2026). A hybrid deep Kronecker and BN-eLU-AlexNet framework for early prediction of learner outcomes in e-learning platforms. Frontiers in Computer Science, 8, Article 1889642. https://doi.org/10.3389/fcomp.2026.1889642

Image Credits: AI Generated

DOI: 10.3389/fcomp.2026.1889642

Keywords: learning analytics, deep learning, e-learning, OULAD, student outcome prediction, DKBNE-AlexNet, hybrid, deep, Kronecker, BN-eLU-AlexNet, framework, early