Psychological Crisis Prediction in College Students Using a Stacked Ensemble of Random Forest, Logistic Regression, and AdaBoost

Abstract

Recently, campus malignant incidents caused by psychological problems among college students have occurred frequently, causing particularly serious impacts on students, families, and schools. At present, the main method for conducting psychological surveys in major universities is through questionnaire surveys. However, this method can make students with mental illness choose answers that are beneficial to them, resulting in misjudgments by psychologists. In response to this issue, this study proposes a stacked fusion model based on Random Forest (RF), Logistic Regression (LR), and AdaBoost, aiming to improve the accuracy and robustness of psychological crisis detection. A dataset consisting of 200 college students' psychological assessment results, behavior logs, and demographic features (totaling 21 indicators) is collected through game simulation and questionnaire surveys. The model collects student information through game simulation, establishes students' psychological files, and analyzes the files to determine whether students have a psychological crisis. The experiment demonstrated that the average absolute error values of LR + iterative algorithm, RF + iterative algorithm, RF + LR, and mixed models were 0.70, 0.69, 0.68, and 0.67, respectively. The root mean square error values were 0.889, 0.885, 0.881, and 0.879, respectively. When the training set was small, the performance of the hybrid model could reach a higher level, and when the validation set was too large, its performance did not significantly decrease. Among all models, the proposed RLA-Stacking model achieved the highest F1-score of 81.8%, outperforming single models and other fusion variants. The research demonstrates that the proposed hybrid model possesses good performance in forecasting psychological crises among college students.