Jui-long Hung

Using hyperscanning measures of brain‑to‑brain synchrony (BBS) with 36 undergraduates, the paper examines online collaborative problem solving (CPS). BBS was higher during problem understanding than solving, and BBS during solving correlated with task performance. Groups with all high CPS students had the highest BBS; some mixed groups reached similar levels. Implications for CPS design are discussed.

Introduces time and location entropy to quantify ‘anytime, anywhere’ behavior using metadata from 5,293 students. High time‑and‑location entropy was associated with lower success; students studying at consistent times and fewer locations—particularly females in this cohort—showed higher performance. Results nuance assumptions about flexibility and success in online learning.

Proposes one‑channel and three‑channel ‘learning image’ representations that transform student course involvement into images for early‑warning prediction. Across experiments with 5,235 students and up to 1,728 variables, CNN approaches achieved higher mid‑semester recall of at‑risk students than SVM, Random Forest, and DNN baselines and enabled subtype identification with interpretable visual cues.

Combining behavioral logs (12,869 students; 14.95M events) with discussion text, the study builds early‑warning models using machine learning and deep learning. Deep models outperformed traditional ML, capturing 51% of at‑risk students by week 8 at 86.8% accuracy; adding text improved recall and accuracy. Linguistic features (e.g., analytic vs. authentic words) related to success and risk.

To address gaps in performance prediction research, the paper proposes a multistage predictive modeling method emphasizing relative engagement. Using datasets from higher education and a K‑12 online school (13,368 students across 300+ courses), the approach outperformed traditional models on accuracy and sensitivity and identified two generalizable engagement predictors across instruction‑ and discussion‑intensive courses.

Focusing on early‑warning for at‑risk online learners, the study modeled student interaction data from LMS logs as indicators of social presence. It tested the hypothesis of a holiday‑effect and compared frequency‑based interaction indicators, showing that interaction frequency is a better predictor of at‑risk status than aggregate amounts, informing timely instructional interventions.

This study modeled early warning indicators of at‑risk students using learning‑management‑system interaction data. Analyses explored whether a holiday‑period effect contributes to failure and compared frequency of interaction versus amount of interaction as predictors. Findings suggest frequency of interaction is a preferable indicator for identifying students needing support.

Introduces time‑series clustering to identify at‑risk online students earlier and with greater accuracy than frequency‑aggregation approaches. In a case study, the best model began capturing at‑risk students by week 10 and revealed ‘holiday effects’ in behavior trajectories, supporting targeted instructional interventions.

Applies text mining to abstracts of 2,997 EDTECH articles (2000–2010) across six SSCI‑indexed journals, identifying 19 research clusters and analyzing productivity by country and journal. The review characterizes rising, stable, and low‑attention topics, and discusses emphases by journal and national strengths through a Critical Theory of Technology lens.

This study investigated longitudinal trends in mobile learning using text mining on 119 refereed journal and proceedings papers (2003–2008). Abstracts were clustered into 12 topic areas across four domains. Results showed growth from 8 to 36 ML articles per year, a dominance of effectiveness/evaluation/personalized systems, and Taiwan’s prominence in several clusters. Findings suggest ML research was in an Early Adopters stage and forecast greater work on strategies and frameworks.

Using text mining of 689 SCI/SSCI-indexed e‑learning publications (2000–2008), this study grouped work into two domains with four groups and 15 clusters, and analyzed subject areas, prolific countries, and journals. It concludes e‑learning research had moved into an early‑majority stage with a shift from ‘does it work?’ to teaching and learning practices, and that national policies shape emphases across countries.

Demonstrates an innovative evaluation approach by combining 23.85 million learning logs, demographics, and end‑of‑course surveys for 7,539 students across 883 K‑12 online courses. Clustering revealed shared characteristics and decision trees predicted course performance and satisfaction. Results show how educational data mining can deepen program evaluation and decision‑making for K‑12 online learning.

Proposes a generic Educational Data Mining (EDM) model for studies of online teaching and learning and demonstrates its procedures via a case study. The model supports pattern discovery and performance prediction to inform instructional improvement and adaptive interventions in online environments.

Using clustering, association rules, and decision trees on LMS data from an experimental study, the paper explains why teacher‑moderated groups outperformed peer‑moderated groups. Teacher presence fostered multi‑source interaction (content, peers, teacher), whereas peer‑moderated groups showed lower participation and relied mostly on student‑content interaction.

Applying data mining to 17,934 LMS server logs from 98 students in an online business course, the study identified behavior patterns, distinguished active vs. passive learners, and extracted predictors of performance. Decision trees and clustering illustrated how mining can support adaptive instruction, early identification of at‑risk learners, and course design improvements.