Modeling Learner Profiles Through Unsupervised HEXAD‐Based Interaction Analysis

ABSTRACT

Learner profiling is a key component of adaptive and personalized learning environments. While the HEXAD framework provides a theoretically grounded taxonomy of motivational user types, its practical use in learning management systems (LMS) typically relies on self‐reported questionnaires, which limits scalability and ecological validity. This study investigates whether meaningful learner profiles inspired by HEXAD can be derived directly from behavioral interaction data in a fully unsupervised manner. Using Moodle log data collected across multiple courses, learner activity is aggregated at the student level and represented through two alternative feature encodings: raw event frequencies and TF–IDF weighting. Hierarchical clustering is applied under consistent conditions, and clustering solutions are evaluated in terms of stability under subsampling, internal diagnostics for cluster resolution, and sensitivity to methodological choices. A cluster configuration with is examined as a theory‐informed reference aligned with the HEXAD framework. Results show that TF–IDF representations produce more stable and robust clustering structures than raw frequencies, particularly at the HEXAD‐aligned resolution. To support interpretability, a dictionary‐based surrogate profiling approach is used to characterise behavioral patterns associated with each cluster, without treating these surrogate labels as ground truth. The analysis reveals imbalances in profile prevalence and highlights the importance of representation choices in unsupervised learner modeling. These findings show that interpretable learner profiles can be derived retrospectively from naturally occurring LMS data without relying on questionnaires. They provide a methodological basis for behavior‐based learner modeling, while the feasibility of generating sufficiently stable profiles early enough for time‐sensitive adaptation remains to be evaluated.