Scholarlink

<p>This study presents a reproducible microgenetic analytics pipeline for decoding fine-grained learning states in virtual reality (VR)–based safety training. Using high-frequency behavioral telemetry from 24 undergraduate civil engineering students in an 18-minute VR heat-stress scenario, we infer latent learner states at 15-second intervals by combining sliding-window segmentation, expert-coded behavioral taxonomies, a Bayesian Cognitive Diagnostic Model (CDM), and KL-regularized feature selection. Learning was modeled as transitions among four microgenetic states—Novice, Developing, Competent, and Expert—grounded in an expert-validated Q-matrix. Across 502 windows, learners spent most of their time in advanced phases (Competent = 35.7%, Expert = 26.9%), yet substantial individual heterogeneity emerged, with some participants remaining novice-persistent and others expert-dominant. Markov chain analysis revealed highly stable learning phases with minimal cross-state transitions, suggesting that effective VR safety performance reflects sustained engagement within coherent microgenetic states rather than frequent oscillation. Under leave-one-subject-out validation, our diagnostic model achieved robust generalization (accuracy = 0.857, macro-F1 = 0.789), with novice-like help-seeking frequency, expert-like proactive behaviors, and temporal efficiency emerging as the most predictive indicators. These findings highlight the limitations of aggregate, post-hoc performance metrics, demonstrate the feasibility of interpretable, probabilistic state estimation in small-N, big-T VR datasets, and point to concrete opportunities for adaptive scaffolding and archetype-informed instructional design in immersive safety training.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Decoding immersive learning States: A reproducible microgenetic pipeline for behavioral data in virtual reality safety training

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Jewoong Moon,
  • Idowu Awoyemi,
  • Stephen Abu,
  • Raissa Marchiori,
  • Siyuan Song

摘要

This study presents a reproducible microgenetic analytics pipeline for decoding fine-grained learning states in virtual reality (VR)–based safety training. Using high-frequency behavioral telemetry from 24 undergraduate civil engineering students in an 18-minute VR heat-stress scenario, we infer latent learner states at 15-second intervals by combining sliding-window segmentation, expert-coded behavioral taxonomies, a Bayesian Cognitive Diagnostic Model (CDM), and KL-regularized feature selection. Learning was modeled as transitions among four microgenetic states—Novice, Developing, Competent, and Expert—grounded in an expert-validated Q-matrix. Across 502 windows, learners spent most of their time in advanced phases (Competent = 35.7%, Expert = 26.9%), yet substantial individual heterogeneity emerged, with some participants remaining novice-persistent and others expert-dominant. Markov chain analysis revealed highly stable learning phases with minimal cross-state transitions, suggesting that effective VR safety performance reflects sustained engagement within coherent microgenetic states rather than frequent oscillation. Under leave-one-subject-out validation, our diagnostic model achieved robust generalization (accuracy = 0.857, macro-F1 = 0.789), with novice-like help-seeking frequency, expert-like proactive behaviors, and temporal efficiency emerging as the most predictive indicators. These findings highlight the limitations of aggregate, post-hoc performance metrics, demonstrate the feasibility of interpretable, probabilistic state estimation in small-N, big-T VR datasets, and point to concrete opportunities for adaptive scaffolding and archetype-informed instructional design in immersive safety training.