JICS_Final_Version

Abstract

Stress impacts students in a great number of ways, influencing not only how they feel but also their in-class performance and the management of day-to-day responsibilities. Actually measuring it, though, is difficult because most measures rely on self-reporting by students, which often lacks reliability. In this work, we investigate whether machine-learning methods can contribute to better stress level predictions based on data from a short online survey. The online survey consisted of basic demographic information, a few questions related to lifestyle, and all 21 items of the University Stress Scale (USS). It yielded a total of 110 completed responses. Four widely used classification models, namely Logistic Regression, Support Vector Machine (SVM), Random Forest, and XGBoost, were trained on three feature combinations to investigate the individual and combined contributions of psychological and non-psychological variables to the prediction. Their performance was estimated using Accuracy, Weighted F1-Score, and Matthews Correlation Coefficient (MCC). In all the experiments, the greatest contribution towards raising the quality of the prediction was made by the psychological features, while among the different models, the best results were obtained using the XGBoost model. We have also compared the self-reported stress of students with their stress categories computed from the USS and found significant mismatches between the two. These point to the suspicion that personal judgment may not reflect the actual stress pattern. These results indicate that a combination of structured psychological scales and machine-learning methods may provide a more reliable approach to understanding student stress.