Machine learning reads VR learning traces to model how students self-regulate

Researchers in Jakarta have shown that the digital footprints children leave inside a virtual reality learning environment can be used to estimate how well they regulate their own learning. Their best model, a Random Forest regressor, predicted students' self-regulated learning scores with a mean absolute error of 0.67 and explained roughly 85 percent of the variance in the scores, far outperforming a simple baseline that always predicted the average. The work, published open access in Discover Artificial Intelligence, is notable less for raw accuracy than for its insistence that every predictive feature be traceable back to an established theory of how learners plan, monitor, and reflect.

Following Zimmerman's cyclical model, researchers treat it as a recurring loop of three phases: forethought or planning, in which learners set goals and select strategies; performance-phase monitoring, in which they track their own progress against those goals; and reflection, in which they evaluate outcomes and adjust future approaches. In digital environments, where learners navigate content independently, SRL becomes a central mechanism that determines whether technology-mediated learning succeeds. Students with stronger self-regulatory skills tend to persist longer, engage more deeply, and construct knowledge more effectively.

Virtual reality complicates this picture in intriguing ways. Immersive environments can foster active participation and experiential understanding by placing learners inside interactive, context-rich scenarios. But the study's authors, Nurzengky Ibrahim, Desy Safitri, Arita Marini, Sujarwo Sujarwo, and Asep Marfu, emphasize that immersion does not automatically produce reflective or goal-oriented behavior.

Learners can become cognitively overloaded or wander through exploratory content without ever planning or evaluating. Whether VR-integrated instruction actually fosters self-regulation has therefore remained an open empirical question, one that survey questionnaires alone struggle to answer because they capture perceptions rather than the moment-to-moment behaviors through which regulation unfolds. The research team's answer was to combine behavioral trace data with explainable machine learning.

One hundred thirty-eight students in Grades 4 through 6, drawn from public and private schools in Jakarta through purposive sampling, completed four structured sessions of roughly 45 minutes each using an environmental education module. The application was built in Unity 3D with the SteamVR plugin and delivered through Oculus Quest 2 standalone headsets. Each session followed a consistent sequence: immersive exploration of virtual environments, guided inquiry tasks, formative assessment with embedded feedback, and reflection activities in which students reviewed their understanding.

Teachers facilitated rather than lectured, handling device orientation and keeping all students on the same instructional timeline. Two data streams were collected. The first consisted of interaction logs recording how students moved through the module: how often they revisited content, how long they spent reviewing material after tasks, how frequently they accessed feedback, whether they revised responses, when they first interacted with the environment, and how many goal-setting clicks they made.

The second was an 18-item SRL questionnaire, adapted for children aged roughly 10 to 12 and administered after the four sessions, with six items each for planning, monitoring, and reflection on five-point Likert scales. The methodological heart of the study lies in its theory-guided feature engineering. Rather than dumping raw clickstream variables into an algorithm and letting the model find whatever patterns exist, the researchers mapped each interaction feature to an SRL dimension before any modeling took place.

Content revisits and feedback access were aligned with monitoring; post-task review duration and response revisions with reflection; early interaction time and initial goal-setting clicks with planning. This mapping, fixed in advance, is what allows the model's later feature importance rankings to be read as statements about learning theory rather than as opaque statistical artifacts. The team is explicit that this alignment is interpretive, and that behavioral proxies cannot fully capture internal cognitive and metacognitive processes.

For prediction, the researchers treated SRL as a continuous outcome and compared two regression algorithms, Random Forest and Support Vector Regression with a radial basis function kernel, against a mean predictor baseline. Data were split 80/20 for training and testing, hyperparameters were tuned with grid search and five-fold cross-validation on the training data alone, and performance was assessed with mean absolute error, root mean square error, and the coefficient of determination. Random Forest won decisively: MAE of 0.67, RMSE of 0.89, and an R-squared of 0.85, compared with 0.78, 1.02, and 0.81 for the support vector model and no explanatory power at all for the baseline.

Cross-validation confirmed stability, with Random Forest's R-squared varying only between 0.83 and 0.87 across folds. The feature importance analysis is where the findings become most interesting for learning scientists. Content revisits carried the highest importance score at 0.187, followed by post-task review duration at 0.165 and feedback access frequency at 0.142, with response revisions close behind at 0.128.

Together, the five highest-ranked features accounted for roughly 72.7 percent of cumulative importance. In contrast, planning-related indicators such as early interaction time (0.105) and initial goal-setting clicks (0.072) contributed far less, and raw time-on-task per session mattered least of all at 0.054. The pattern suggests that monitoring- and reflection-related behaviors leave richer, more detectable traces in immersive environments than planning behaviors do, echoing earlier arguments that forethought processes are theoretically central but behaviorally elusive in log data.

The authors are unusually candid about their study's central weakness. When participants were grouped descriptively by SRL level, 94.7 percent fell into the High category, 5.3 percent into Medium, and none into Low. This severe imbalance likely biased the models toward the dominant group, potentially inflating accuracy and R-squared while leaving the model poorly equipped to distinguish learners at lower levels of self-regulation.

The team recommends that future work employ balanced samples or class-balancing techniques such as oversampling, class weighting, or synthetic data generation before such models are trusted across the full spectrum of learners. Even with those caveats, the study offers a template for how immersive learning analytics might develop. The correlations between behavior and reported SRL were substantial, with post-task review duration at r = 0.73 and content revisits at r = 0.71, but the researchers stress these are predictive associations, not causal effects, and that feature importance reflects a model's internal structure rather than psychological mechanism.

Their broader argument is that machine learning in education earns its keep not through prediction alone but through interpretable evidence that complements existing theory. Trace data, on this view, offers a partial but valuable window into self-regulation, one that should be read alongside self-reports and observation. For instructional designers, the practical implication is sobering: immersive engagement by itself does not guarantee self-regulation, so VR modules need explicit scaffolds for planning and reflection rather than relying on the technology's affordances to do the work.

Subject of Research: Explainable machine learning modeling of elementary students' self-regulated learning from VR e-module interaction data

Article Title: Modeling students’ self-regulated learning from virtual reality-integrated e-module interaction data

Article References: Ibrahim, N., Safitri, D., Marini, A., Sujarwo, S., & Marfu, A. (2026). Modeling students’ self-regulated learning from virtual reality-integrated e-module interaction data. Discover Artificial Intelligence, 6(1), Article 986. https://doi.org/10.1007/s44163-026-02103-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02103-4

Keywords: self-regulated learning, virtual reality, e-modules, machine learning, learning analytics, Random Forest, feature engineering, educational technology, elementary education, behavioral trace data, explainable AI, Indonesia