High predictive accuracy in machine learning models trained on digital platform activity tracks engineered proxy definitions rather than true psychological profiles, according to a study published in the International Journal of Computational Intelligence Systems. Researchers Ricardo Tesoriero, Felipe Costa-Tebar, Jose A. Gallud, and Jose A. Gamez examined how behavioral proxies operate when direct user assessment is unavailable. Their investigation demonstrates that high machine learning performance reflects the recovery of engineered behavioral proxies rather than the independent psychometric identification of latent motivational profiles.
Digital platforms routinely record detailed interaction logs to facilitate behavior-based user modeling when direct psychometric testing cannot be administered. However, when both the target proxy labels and the predictive input features are derived from the exact same underlying behavioral traces, measured model performance heavily depends on how researchers construct, represent, and encode the target. To investigate these fundamental dependencies, the research team analyzed interaction traces collected from 1,037 Moodle users, encompassing a total of 109,729 mapped interaction events within an educational platform environment.
The investigators developed and compared two distinct proxy-label constructions designed to infer HEXAD-inspired user profiles from the recorded interaction events. One proxy-label formulation relied on raw event frequencies, while the alternative approach utilized term frequency–inverse document frequency weighting. When applied to the cohort of 1,037 users, the two construction methods assigned divergent profile labels to 136 individuals, representing 13.11 percent of the user population. In addition to reclassifying over thirteen percent of the participants, the two weighting strategies produced noticeably different overall class distributions across the dataset.
To evaluate predictive behavior, the authors paired the proxy constructions with corresponding feature representations and evaluated five supervised learning algorithms across multiple testing conditions. Their testing pipeline incorporated optional in-fold Synthetic Minority Over-sampling Technique, commonly known as SMOTE, to balance classes. The evaluation protocol employed repeated stratified five-fold cross-validation conducted over ten distinct repeats. Across matched experimental configurations, altering solely the method of proxy construction generated observed differences in macro-averaged F1 scores of up to 0.0786 across the tested models.
Despite these observed variations in macro-averaged F1 scores, none of the pairwise performance comparisons remained statistically significant following the application of Holm correction. The highest overall performance was achieved by Extreme Gradient Boosting using term frequency–inverse document frequency features, raw proxy labels, and in-fold oversampling. This specific configuration attained a mean macro-averaged F1 score of 0.9147 alongside a mean balanced accuracy of 0.9610. However, after applying statistical Holm corrections, the comparisons did not establish a definitive performance advantage for Extreme Gradient Boosting over Random Forest.
A deterministic analysis conducted by the researchers revealed perfect recovery of term frequency–inverse document frequency proxy labels. The analysis also demonstrated perfect mathematical recovery of raw proxy labels for users who exhibited a unique raw maximum. These analytical findings underscore that the initial proxy-label construction fundamentally defines the supervised learning task itself. Consequently, the authors conclude that strong predictive metrics indicate successful automated recovery of engineered behavioral proxies instead of genuine, independent psychometric classification of underlying user motivation.
Subject of Research: Machine learning inference of HEXAD user profiles from interaction traces
Article Title: Behaviour-based Inference of HEXAD-inspired User Profiles from Interaction Traces: The Role of Proxy Label Construction
Article References: Tesoriero, R., Costa-Tebar, F., Gallud, J. A., & Gamez, J. A. (2026). Behaviour-based Inference of HEXAD-inspired User Profiles from Interaction Traces: The Role of Proxy Label Construction. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01653-w
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01653-w
Keywords: HEXAD user types, interaction traces, proxy-label construction, user profiling, Moodle interaction logs, TF-IDF weighting, raw event frequency, supervised learning, Extreme Gradient Boosting, Random Forest, SMOTE oversampling, stratified cross-validation, macro-averaged F1 score, balanced accuracy, Holm correction, behavioral user modeling, deterministic recovery
Cite Scienmag News
Denise Maddox. (November 19, 2026). Behavioral Profiling From Digital Logs Reflects Proxy Construction Rather Than Motivation. Scienmag. https://scienmag.com/behavioral-profiling-from-digital-logs-reflects-proxy-construction-rather-than-motivation/
Denise Maddox. "Behavioral Profiling From Digital Logs Reflects Proxy Construction Rather Than Motivation." Scienmag, 19 November 2026, https://scienmag.com/behavioral-profiling-from-digital-logs-reflects-proxy-construction-rather-than-motivation/. Accessed 19 November 2026.
Denise Maddox. "Behavioral Profiling From Digital Logs Reflects Proxy Construction Rather Than Motivation." Scienmag. November 19, 2026. https://scienmag.com/behavioral-profiling-from-digital-logs-reflects-proxy-construction-rather-than-motivation/








