Behaviour-based Inference of HEXAD-inspired User Profiles from Interaction Traces: The Role of Proxy Label Construction

Abstract Digital platforms generate interaction logs that support behaviour-based user modelling when direct assessment is unavailable. However, when proxy labels and predictive features are derived from the same behavioural source, measured performance may depend on how the target is constructed, represented, and encoded in the predictors. This study examines these dependencies using HEXAD-inspired behavioural proxies derived from 1,037 Moodle users and 109,729 mapped interaction events. Two proxy-label constructions based on raw event frequencies and term frequency–inverse document frequency weighting were evaluated with corresponding feature representations, five supervised learning algorithms, optional in-fold Synthetic Minority Over-sampling Technique, and repeated stratified five-fold cross-validation with ten repeats. The proxy constructions assigned different labels to 136 users (13.11%) and produced different class distributions. Across matched configurations, changing only the proxy construction yielded observed macro-averaged F1 score differences of up to 0.0786, but no comparison remained statistically significant after Holm correction. The highest observed performance was obtained by Extreme Gradient Boosting using term frequency–inverse document frequency features, raw proxy labels, and in-fold oversampling, reaching a mean macro-averaged F1 score of 0.9147 and a mean balanced accuracy of 0.9610. Corrected comparisons did not establish an advantage over Random Forest. Deterministic analysis showed perfect recovery of term frequency–inverse document frequency proxy labels and perfect recovery of raw proxy labels for users with a unique raw maximum. These findings show that proxy-label construction defines the supervised task and that high predictive performance should be interpreted as recovery of engineered behavioural proxies rather than independent psychometric identification of latent motivational profiles.