Comparative Analysis Of Ensemble Learning Algorithms For Student Mental Health Classification: Balancing Predictive Performance And Computational Efficiency

Student mental health disorders have become an increasingly important issue due to their significant impact on academic performance and psychological well-being. Machine learning-based approaches offer a promising alternative for supporting early detection; however, only a limited number of studies have conducted comprehensive evaluations of various ensemble learning algorithms by simultaneously considering predictive performance, computational efficiency, and statistical significance. This study aims to compare ten classification algorithms, namely Random Forest, Extra Trees, Gradient Boosting, Hist Gradient Boosting, AdaBoost, XGBoost, LightGBM, CatBoost, Logistic Regression, and Stacking Ensemble, using 5-fold Stratified Cross-Validation. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, Receiver Operating Characteristic–Area Under the Curve (ROC-AUC), training time, inference time, confusion matrix, feature importance, as well as One-Way Analysis of Variance (ANOVA) and post-hoc Tukey’s Honestly Significant Difference (Tukey HSD) tests. The experimental results show that CatBoost achieved the highest classification performance, with an Accuracy of 84.86%, an F1-Score of 87.19%, a Recall of 88.03%, and an ROC-AUC of 0.9220. XGBoost delivered comparable predictive performance with a substantially shorter training time (3.92 seconds), while LightGBM was the most computationally efficient algorithm, requiring an average training time of only 0.85 seconds. In contrast, the Stacking Ensemble required the longest training time (215.41 seconds) without providing a statistically meaningful performance improvement over the boosting-based models. The ANOVA results revealed significant differences in ROC-AUC among the evaluated models (p < 0.001), whereas the Tukey HSD test indicated that most boosting algorithms did not differ significantly from one another. These findings provide practical guidance for selecting classification algorithms by balancing predictive performance and computational efficiency to support early mental health screening systems for university students.