Construction of an Early Warning Model for Academic Burnout among High School Students Based on Learning Behaviour Data

Abstract

The issue of academic burnout among adolescents has gained increased prominence in the field of educational psychology; nevertheless, current measurement instruments and predictive models remain limited due to their use of cross-sectional self-reports and black box modeling pipelines which do not allow for any actionable solutions in the school environment. This paper suggests an explainable machine-learning approach to identifying cases of academic burnout among high school students based on their observable learning behaviors. More specifically, by employing a publicly accessible dataset that includes information about 649 adolescents and contains no validated burnout measurement, and considering the multi-dimensional construct of burnout, the modeling pipeline involves a composite proxy label which implies converging data on at least two aspects of burnout, predictors which consist of observable behaviors, three supervised learning algorithms (logistic regression, random forest, and extreme gradient boosting), and the SHapley Additive exPlanations method which is implemented at both population and individual levels. Crucially, all variables used to construct the proxy label, together with any feature derived from them, were excluded from the predictor matrix to prevent target leakage. Once leakage was removed, the models performed only modestly (held-out ROC-AUC ≈ 0.74 and PR-AUC ≈ 0.52–0.56 for the tree-based models), far below the near-ceiling values obtained when label-defining variables are retained, indicating that such high performance is largely circular. The strongest leakage-free predictors were early-term grades, alcohol use, free time, and age rather than attendance and study time. Thus, observable behaviors carry only a weak-to-moderate, non-causal signal for a proxy burnout label, and a deployable screening tool would require an independent dataset containing a measured burnout instrument such as the MBI-SS.