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

1 Introduction

Recently, the high-quality online education demand has increased owing to the evolving learning preferences and growth of digital platforms. As reliance on learning management systems (LMS) has increased, maintaining learner outcome prediction has become challenging and important. As network technologies, such as fifth generation (5G) mobile networks and high-speed networks, have achieved great development; technologies such as artificial intelligence (AI) and big data have become integral parts of human life. Due to the rapid growth of network technologies, online learning has had a major impact on the educational system (Li, 2024). The COVID-19 pandemic has caused varied degrees of impact on economies throughout the world, and most sectors have shifted their administrative procedures to avoid loss. Owing to the effect of national-level anti-epidemic regulations, online teaching has emerged as the most optimal method for education (Ye, 2020). There has been a lack of appropriate quality control processes for online courses, and hybrid teaching reform has lacked logical and operational enhancement (Zhang, 2021). The rising demand for online course quality is compelled to begin with the assessment of courses through learners and educators, and then it evaluates course quality. Enhancing the quality of online education is a critical topic that cannot be overlooked as higher education advances.

Online teaching techniques offer many features that are more explicit in the presentation of educational objectives that are capable of face-to-face interaction using visual sensors (Almubarak et al., 2025). Compared with traditional offline teaching modes, online teaching may achieve superior teaching with the support of a teaching environment. The benefit of online teaching is that it may be done using inexpensive electronic equipment for both learners and instructors (Maatuk et al., 2022; Frederick and Van Gelderen, 2021). Online learning and instruction are performed using computer vision, and the interaction between humans and computer technologies has significantly enhanced teaching and learning capabilities (Bruggeman et al., 2022). It motivated students and teachers to join the classes. Several academics have also conducted studies to improve the efficiency of online education and to improve the online teaching approach to that of in-person instruction (Nikou, 2021). Educational institutions have implemented extensive hybrid teaching improvements to conventional and online courses to fully realize the impact of excellent educational materials and satisfy the increasingly individualized learning needs of learners. An online education evaluation system was established based on the analytical hierarchy process. The importance of the sub systems and each indicator was examined, and the risk values of each indicator were established using the most recent practical experience available (Ma, 2024).

Many researchers have shifted their focus to the evaluation of teaching performance, while others have focused on understanding students’ performance in the online teaching learning environment (Kang and Park, 2022; Han et al., 2021). The assessment process used to evaluate online learning in higher education institutions considers academic targets while keeping up with ongoing advancements (Huang, 2017). Universities across the world should improve the management and implementation of strategies, such as audit, evaluation, and monitoring, and quality standards, to guarantee standard online courses (Cowin and Moroney, 2018). To develop effective educational objectives, it is required to enhance the educational system, and instructors need to establish a system to evaluate the learning outcomes (Moreno-Murcia et al., 2015). Various courses should develop independent assessment methodologies, alternative human–machine models, and two-way indicators for the assessment of teachers and students (Poquet et al., 2022).

In this study, the prediction task was operationalized as the early prediction of learner assessment outcomes rather than direct evaluation of instructor teaching quality. We used course-design attributes, learner characteristics, and early engagement traces from the Open University Learning Analytics Dataset (OULAD) to predict a weighted continuous assessment score. The proposed DKBNE-AlexNet framework combines a parameter-efficient Deep Kronecker Network with a feature-token BN-eLU convolutional branch and gated residual fusion. The study was designed to test generalization to previously unseen module presentations through presentation-grouped repeated cross-validation.

1.1 Aim and research questions

This study predicts learner assessment outcomes rather than directly measuring or evaluating the quality of teaching. This study addressed the following research questions:

RQ1. To what extent can learner assessment outcomes be predicted early from course design attributes and learner engagement data available during the first 25% of a module presentation?

RQ2. How does DKBNE-AlexNet perform relative to Random Forest and LightGBM when all models are evaluated using identical repeated presentation-grouped cross-validation?

RQ3. What predictive contribution is provided by the DKN and BN-eLU-AlexNet components of the proposed hybrid architecture?

The paper is organized into seven sections. Section 2 discusses the motivation and presents a literature review of online learning quality prediction. Section 3 describes the OULAD study design, cohort construction, target, and feature engineering. Section 4 presents the DKBNE-AlexNet-enabled architecture. Section 5 describes the experimental protocol. Section 6 reports the comparative results, and Section 7 discusses the findings, educational implications, and Section 8 concludes the study.

2 Motivation

Online education provides broader access and flexibility to learners and educators; however, all digital platforms do not provide the same good experience. This inconsistency intends the need for a robust and data-driven approach to improve online teaching learning experience. Here, teachers can opt for advanced methods to provide feedback on students' learning and performance.

2.1 Literature survey

Ma (2024) compared various algorithms, such as DT, RF, LR, SVM, and LSTM, for grading learning outcomes using PCA based feature selection. The LSTM RNN-RF model demonstrated the highest accuracy and adaptability in student performance prediction, achieving an accuracy of 98.5%. For the feature selection, has used principal component analysis (PCA) across six evaluation parameters and tested six different multi-classification models. The practical educational challenges of low accuracy and low algorithmic adaptability were addressed directly.

Li (2024) used a machine learning-based approach for teaching evaluation with clear metric definitions. An interactive teaching evaluation system was developed using machine learning algorithms (SVM, KNN, Naive Bayes, and logistic regression). The SVM classifier significantly outperformed the other classifiers, ranking as SVM > LR > KNN > NB in classification effectiveness. The SVM achieved 92.1% accuracy, indicating the model’s suitability for developing interactive teaching evaluation systems under emotional education frameworks. Hou et al. (2022) developed a multi-task cascaded convolutional neural network-based learning outcome analysis. This model did not include learning outcome assessment with computer vision-based attention estimation of students.

Hou (2021) devised a teaching quality prediction scheme using an adaptive BP neural network. The model showed high reliability while evaluating the standard of education offered by higher education institutions, as well as excellent efficacy for handling large datasets. However, this model requires regular modification of relevant crucial variables to increase the processing power to evaluate instruction quality.

Wang (2022) developed a convolutional neural network (CNN)-based online learning quality analysis. The flexibility and correlation processing capacity of the model were improved without increasing the computing cost. However, weak texture problems, multi-target movement, and single-frame image optical flow were not detected in the data during the quality assessment. Liu (2022) devised a separate long-term recurrent convolutional network (SLRCN) model for online teaching quality assessment. When assessing the quality of instruction, the model offered a simple design and low requirement for data preprocessing and reduced the probability of overfitting issues. However, it was unable to examine the associated connection among the changes in emotion throughout the learning and training states. Almubarak et al. (2025) developed a computer-vision framework using three state-of-the-art object detectors–YOLOv8, Faster R-CNN, and RetinaNet–to identify eleven classroom interaction categories. The study evaluated the performance of a labeled dataset of 7,259 images collected from real classroom settings. YOLOv8 achieved the best performance with a mean Average Precision (mAP) of 85.8%, while Faster R-CNN and RetinaNet performed competitively but were outperformed. The authors concluded that deep learning-based detection provides more reliable insights into teacher–student interactions compared to traditional approaches. This conclusion is supported by evidence-based evaluation and improved feedback. Xuan Lam et al. (2024) used 13 features derived from student engagement patterns, progress tracking, and time spent on assignments. The research compared five machine algorithms: random forest (RF), support vector machine (SVM), logistic regression (LR), naive Bayes (NB), and k-nearest neighbor (KNN). The findings from the research was that if classifiers are stacked accurately, they predict students' learning outcomes effectively. This study did not provide specific accuracy metrics for educational assessment. The emphasis of this research is that integration of data driven forecasting helps to improve the teaching methodologies and enable a proactive learning environment.

Previous studies often focus on evaluating teaching quality, whereas the present study focuses specifically on the early prediction of learner outcomes using Open University Learning Analytics Dataset (OULAD) behavioral and course-design data.

2.2 Challenges

The challenges faced by different techniques to predict online learning outcomes as follows

• Ma (2024) used machine learning, and the network structure was modified to increase the reliability of the quality monitoring system. However, the convergence speed of the model was extremely slow throughout the quality assessment.
• During the safety evaluation of online teaching quality, the AHP and PSO-BP models employed by Li (2024) successfully monitored the network’s historic optimum weights and solved network faults. However, algorithmic strategies were not included to maximize the threshold-modification issues.
• The MTCNN model used by Hou et al. (2022) was more accurate in assessing student status. Nevertheless, the system operated slowly because it had difficulty detecting small targets. Furthermore, the accuracy of the model decreased because the expression recognition network was not trained with the matching student classroom face dataset.
• The CNN model utilized by Wang (2022) showed quick computation while assessing educational quality; however, it was unable to handle serious issues with occlusion and displacement during the assessment process.

3 Data preparation and study design

This study uses the Open University Learning Analytics Dataset (OULAD), a publicly available dataset. This dataset contains learners' demographics, registrations, module metadata, assessment records, learning resources, and interaction logs (Kuzilek et al., 2017). The analysis unit is an enrolment, which is a combination of module, presentation, and learner.

3.1 Dataset and analytical cohort

The starting OULAD cohort contained 32,593 participants. Enrolments were excluded when a learner had unregistered on or before the early prediction cutoff because the outcome of those cases was already determined by withdrawal before the prediction point. Enrolments were required to have at least one valid non-banked continuous assessment score. The final analytical cohort comprised 24,153 enrolments from 22 module presentations, seven modules, and 22,234 unique learners. The OULAD does not provide a reliable instructor identifier; therefore, the number and individual effects of instructors could not be determined from the available data (Table 1).

Table 1. Analytical cohort and prediction setup.

3.2 Early-prediction window and target definition

Early engagement was restricted to VLE interactions occurring on or before 25% of the module presentation length, corresponding to approximately days 58–67 across presentations. The prediction target was a continuous learner outcome on a 0–100 scale, calculated as the weighted mean of tutor-marked assessments (TMA) and computer-marked assessments (CMA). Therefore, class distribution statistics were not applicable. 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 included as a predictor variable.

3.3 Feature construction and leakage controls

The final feature matrix contained 84 predictors. Learner-level predictors included demographic and registration attributes, as well as 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. The held-out presentation groups constituted the test folds, whereas preprocessing and model fitting were conducted only on the corresponding training data. Course-design predictors described the composition of VLE materials, the number and spacing of assessments, assessment weights, and presentation length. Course and presentation identifiers were excluded from the predictor matrix because they defined the grouped validation units. Missing-value imputation, standardization, and categorical encoding were fitted inside each training fold. All data transformations were estimated exclusively from the training portion of each outer fold and subsequently applied to the corresponding held-out fold.

4 Proposed DKBNE-AlexNet architecture

DKBNE-AlexNet combines a deep Kronecker network branch with a feature-token BN-eLU-AlexNet-inspired branch, a wide linear residual path, and gated residual fusion. The architecture is designed for heterogeneous tabular learning analytics data and, therefore, does not reshape the feature vector into an artificial two-dimensional image.

4.1 Deep Kronecker network branch

The DKN branch offers parameter-efficient representation learning by using Kronecker structured linear layers (Feng and Yang, 2022). For a rank ‘R’ parameterization, a weight matrix ‘W’ is represented as a sum of Kronecker products, W = Σr(Ar ⊗ Br), by reducing the number of free parameters with respect to an unrestricted dense matrix. The final configuration, DKN, maps the processes input through 256 and 128 dimensional Kronecker layers with eLU activation, LayerNorm, dropout, and residual projection to a 96-dimensional latent representation.

4.2 Feature-token BN-eLU-AlexNet branch

AlexNet was originally designed for two-dimensional image data. To avoid imposing arbitrary spatial adjacency on heterogeneous OULAD columns, each scalar predictor was first converted into a feature-specific 24-dimensional token. Five 1 × 1 Conv1D stages with batch normalization and eLU activation then transform feature tokens without mixing neighboring columns because of their position in the input matrix. Attention, mean, and max pooling are combined to form a global representation that is projected onto the fusion space. This branch preserves a five-stage convolutional transformation while adapting to tabular feature tokens. The proposed architecture is AlexNet-inspired rather than a conventional image-based AlexNet (Alaeddine and Jihene, 2021).

4.3 Gated residual fusion and prediction head

The full model is anchored by the DKN prediction. The DKN and convolutional representations are projected to a common space and combined through their direct representations, element-wise interaction, and absolute difference. A learned sigmoid gate scales the residual correction produced by the convolutional/fusion path. A wide linear path from the processed original predictors is added to the final prediction so that simple additive relationships can be represented directly. The model is trained in two stages: first, the DKN anchor is optimized, and then the convolutional/fusion branch is trained as a residual correction while the anchor is frozen. This design allows the inner validation procedure to retain the anchor when an added correction is not beneficial (Table 2).

Table 2. Final DKBNE-AlexNet implementation configuration.

Schematic Figures 1 and 2 were prepared using AI-assisted drafting and then verified against the feature-construction and model-implementation code. All quantitative plots in the Results section were generated directly from the experimental output files.

5 Experimental protocol

5.1 Grouped repeated cross-validation

Generalization was evaluated using five-fold cross-validation grouped by module presentation and repeated three times. All enrolments belonging to a held-out module 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. The final audit confirmed 15 outer evaluations for every model and complete repeated out-of-fold predictions. Five grouped folds were evaluated in each repetition, resulting in 15 held-out evaluations across three repetitions.

5.2 Comparative models

The main comparison included Random Forest and LightGBM as strong tabular machine-learning baselines, together with the two principal neural components–DKN and BN-eLU-AlexNet–and the complete DKBNE-AlexNet model. All models used identical outer grouped partitions.

5.3 Evaluation metrics and statistical analysis

Performance was assessed using the mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²). MAE was used as the primary metric because it is directly interpretable in terms of score points. The uncertainty for the pooled MAE was estimated using a group-aware bootstrap over module presentations with 5,000 resamples. Model comparisons were conducted using paired Wilcoxon signed rank tests on representation level MAE values and Holm adjustment for multiple comparisons. Held-out group-aware permutation importance was evaluated to assess model interpretability by measuring the performance degradation in MAE following the disruption of individual predictors while keeping the trained parameters constant.

6 Results

6.1 Comparative predictive performance

Random Forest achieved the lowest pooled MAE (10.916), followed closely by LightGBM (10.918). BN-eLU-AlexNet achieved an MAE of 11.173, whereas DKN and DKBNE-AlexNet achieved MAEs of 11.416 and 11.424, respectively. The proposed model obtained an RMSE of 14.843 and an R² of 0.117. Random Forest and LightGBM have lower point-estimate MAEs, while the difference between DKBNE-AlexNet and Random Forest was not statistically significant (Figures 3, 4; Table 3).

Table 3. Predictive performance of selected models.

6.2 Generalization across module presentations

The presentation-level performance varied across the 22 held-out module presentations, as shown in Figure 5. The proposed DKBNE-AlexNet displayed a compact repeated-fold MAE distribution (fold-level MAE SD approximately 0.61), indicating improved stability as per the observed dispersion of fold-level MAE. The presentation-level distributions of the five selected models overlapped considerably, consistent with the non-significant Holm-adjusted pairwise comparisons against the Random Forest (Table 4).

Table 4. Presentation-level paired comparisons against random forest.

6.3 Prediction and residual analysis

The observed-versus-predicted plot in Figure 6 shows that DKBNE-AlexNet produces individualized predictions rather than collapsing to the global target mean. The residual mean was + 0.66 score points, with a residual standard deviation of 14.53. The residual pattern nevertheless indicates compression toward the center of the outcome distribution: very low observed scores tend to be overpredicted, while very high observed scores tend to be underpredicted.

6.4 Component-wise ablation analysis

The component analysis compared BN-eLU-AlexNet alone, DKN alone, and the complete hybrid architecture alone. BN-eLU-AlexNet produced the lowest neural model point estimate (MAE 11.173), followed by DKN (11.416) and complete DKBNE-AlexNet (11.424). The overlapping group-aware confidence intervals indicate that the present experiment did not establish a clear average error advantage from hybrid fusion. Therefore, the result is interpreted as evidence that both branches contain useful predictive information rather than proof that fusion improves accuracy (Figure 7).

6.5 Feature attribution and model interpretation

The held-out group-aware permutation importance identified early engagement and learner-context variables as the most consistent contributors to DKBNE-AlexNet predictions. The number of distinct learning materials accessed had the largest mean importance (ΔMAE 0.329; 95% CI 0.163–0.528; Holm-adjusted p < 0.001), followed by highest previous education (ΔMAE 0.208; 95% CI 0.177–0.238; adjusted p < 0.001), active learning days (ΔMAE 0.168; 95% CI 0.117–0.221; adjusted p < 0.001), forum engagement (ΔMAE 0.103; 95% CI 0.073–0.134; adjusted p < 0.001), and IMD group (ΔMAE 0.086; 95% CI 0.068–0.105; adjusted p < 0.001). The timing of the first recorded interaction also remained significant after Holm correction (ΔMAE 0.025; adjusted p = 0.001). The individual course-design variables generally showed smaller and less stable effects. These importance estimates represent predictive associations rather than causal effects (Figure 8).

7 Discussion

The updated evaluation demonstrates that student performance can be forecast with reasonable precision using behavioral data captured within the initial 25% of a course offering. Decision tree ensembles, specifically random forest and light gradient boosting machine, yielded the most favorable point metrics, underscoring the efficacy of ensemble techniques on varied learning analytics data. Furthermore, DKBNE-AlexNet demonstrated competitive presentation-level metrics without significant differences while maintaining consistent stability across the presentation-grouped partitions. Feature attribution revealed that the diversity and continuity of early VLE interactions offer superior predictive value compared to static course-structure metrics.

7.1 Educational implications

The framework is most appropriately interpreted as an early-warning learning-analytics instrument rather than a direct evaluation of teacher effectiveness. Predictive estimates constructed during the initial 25% of a module presentation can inform targeted instructional interventions, academic support prioritization, or systematic monitoring of presentations exhibiting elevated outcome risks. Because the 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 proposed model can support the early identification of learners who may require additional academic support.

7.2 Limitations

Several limitations should be considered. First, 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; therefore, instructor effects cannot be estimated separately from presentation effects. Second, the analysis covers 22 module presentations from a single distance-learning institution, which limits external generalization. Third, the residual analysis shows regression toward the mean at extreme scores. Fourth, the ablation study does not demonstrate a clear mean-error advantage from combining the DKN and BN-eLU-AlexNet branches. Finally, permutation importance identifies predictive associations but does not establish causality. Future studies could complement permutation-based feature attribution with instance-level explainable AI methods to improve the transparency of individual predictions.

8 Conclusion

This study developed and evaluated DKBNE-AlexNet for the early prediction of learner assessment outcomes from OULAD. For RQ1, the proposed model achieved MAE 11.424, RMSE 14.843, and R² 0.117 under five-fold presentation-grouped cross validation repeated three times. For RQ2, Random Forest and LightGBM achieved slightly lower MAEs of 10.916 and 10.918, respectively; however, the presentation-level difference between the proposed model and Random Forest was not statistically significant. For RQ3, BN-eLU-AlexNet and DKN retained most of the predictive performance, and the current experiment did not establish a clear average-error advantage from their fusion. The feature-attribution analysis showed that the diversity and regularity of early engagement were among the strongest predictive contributors. Overall, the findings support DKBNE-AlexNet as a competitive and stable early-prediction framework, while emphasizing that strong tabular baselines, transparent uncertainty estimation, and careful interpretation are essential in learning analytics research.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. A grammar checker was used to assist with language editing, restructuring, and refinement of manuscript text. Generative AI generation/refinement of schematic figures. OpenAI ChatGPT was used for these purposes. All analyses, numerical results, interpretations, references, and final manuscript content were reviewed and verified by the authors, who take full responsibility for the submitted work.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Konstantinos T. Kotsis, University of Ioannina, Greece

Reviewed by: Liudmyla Gryzun, Simon Kuznets Kharkiv National University of Economics, Ukraine

Taufikin Taufikin, IAIN Kudus, Indonesia