Learning Management Systems (LMS) generate rich activity and interaction logs that can be exploited using machine learning techniques. This study models temporal engagement patterns, such as early, middle, late, weekend, and night activity, derived from Moodle logs in multiple undergraduate courses. It constructs temporal feature vectors per-student, applies k-means clustering to uncover behavioral patterns, and then uses ANOVA and Kruskal–Wallis tests to assess whether patterns differ in final grades. Results show that the predictive value of temporal patterns is highly course-dependent; in some courses, structured early engagement aligns with higher achievement, whereas in others, heavy weekend and night usage is associated with the best outcomes. To complement the obtained quantitative analyses, a Large Language Model (LLM) (i.e., ChatGPT) is evaluated as a zero-shot classifier that receives only natural-language summaries of temporal behavior and predicts performance tiers. While accuracy is limited, the model produces a coherent approach, indicating value as an interpretable layer on top of statistical analysis. The work demonstrates a generalizable pipeline for temporal feature engineering, unsupervised profiling, and LLM-based reasoning over LMS data for early risk detection in digital learning environments.
Shehada et al. (Thu,) studied this question.