TY - JOUR
T1 - Performance prediction in VR co-creation learning environments
T2 - Dual-phase vs. single-phase feature extraction and predictor evaluation in student engagement
AU - Wang, Hsin Yun
AU - Sun, Jerry Chih Yuan
N1 - Publisher Copyright:
© The Author(s) 2026.
PY - 2026/7
Y1 - 2026/7
N2 - This experimental study examined the impact of feature extraction on predicting semester integrated performance assessment (IPA) in virtual reality (VR) co-creation learning environments. A total of 83 students attended an 18-week-long VR co-creation course at a northern Taiwan public high school using a blended online-merge-offline (OMO) model. Evidence-based multimodal data, including log data, assessment scores, electroencephalography (EEG), and demographics, were used for dual-phase and single-phase Random Forest (RF) models which predicted binary pass/fail semester IPA. The dual-phase RF model, with PCA for feature extraction, demonstrated superior accuracy compared to the single-phase model; however, the single-phase model still achieved acceptable performance. Additionally, both RF models achieved high accuracy in predicting outcomes using EEG data, highlighting the potential of integrating physiological responses into educational settings. Further, prior knowledge and age emerged as crucial demographic features, whereas gender showed minimal influence on prediction accuracy. The study indicates limitations such as the scarcity of diverse cognition data, the lack of performance dashboards, and the absence of longitudinal analysis. Future studies should expand cognition data diversity, develop real-time algorithm-driven dashboards, and employ longitudinal studies to explore growth and community dynamics in VR co-creation.
AB - This experimental study examined the impact of feature extraction on predicting semester integrated performance assessment (IPA) in virtual reality (VR) co-creation learning environments. A total of 83 students attended an 18-week-long VR co-creation course at a northern Taiwan public high school using a blended online-merge-offline (OMO) model. Evidence-based multimodal data, including log data, assessment scores, electroencephalography (EEG), and demographics, were used for dual-phase and single-phase Random Forest (RF) models which predicted binary pass/fail semester IPA. The dual-phase RF model, with PCA for feature extraction, demonstrated superior accuracy compared to the single-phase model; however, the single-phase model still achieved acceptable performance. Additionally, both RF models achieved high accuracy in predicting outcomes using EEG data, highlighting the potential of integrating physiological responses into educational settings. Further, prior knowledge and age emerged as crucial demographic features, whereas gender showed minimal influence on prediction accuracy. The study indicates limitations such as the scarcity of diverse cognition data, the lack of performance dashboards, and the absence of longitudinal analysis. Future studies should expand cognition data diversity, develop real-time algorithm-driven dashboards, and employ longitudinal studies to explore growth and community dynamics in VR co-creation.
KW - Feature extraction
KW - Machine learning (ML)
KW - Multimodal data analytics (MMLA)
KW - Predictor evaluation
KW - Random forest (RF)
KW - Virtual reality (VR) co-creation
UR - https://www.scopus.com/pages/publications/105034091852
U2 - 10.1007/s10639-026-13894-5
DO - 10.1007/s10639-026-13894-5
M3 - Article
AN - SCOPUS:105034091852
SN - 1360-2357
VL - 31
SP - 4003
EP - 4039
JO - Education and Information Technologies
JF - Education and Information Technologies
IS - 11
ER -