Predicting Online Course Completion: A Comparative Analysis of Machine Learning Models and the Incremental Role of Structural Learning Features

다국어 초록 (Multilingual Abstract)

This study examines the predictive determinants of online course completion by comparing behavioral engagement variables with structural learning features. As digital learning environments have rapidly expanded following the COVID-19 pandemic, accurately predicting learner persistence has become increasingly important. However, prior research has predominantly relied on activity-based behavioral logs, such as time spent learning and assignment submission frequency, while giving limited attention to structural characteristics of learning environments. Using a large-scale dataset comprising 100,000 learner profiles, this study compares the predictive performance of Logistic Regression, Random Forest, Support Vector Machine, and XGBoost models. In addition, it statistically evaluates the incremental predictive value of structural learning variables, including learning path type and engagement consistency. The results indicate that boosting-based ensemble methods achieved the strongest overall predictive performance, although linear models remained competitive. More importantly, the inclusion of structural learning features significantly improved model performance, demonstrating that course completion is not solely determined by engagement intensity but is also shaped by the structural design of the learning environment. This study contributes theoretically by reframing online learning persistence beyond behavioral activity measures and methodologically by emphasizing the importance of feature design in predictive modeling. Practically, the findings suggest that digital learning platforms should incorporate structural learning indicators into early risk detection systems and instructional design strategies.