SMOTEENN hybrid sampling combined with stacking ensemble model achieves optimal performance in predicting depression tendency among college students: a comparison of 84 models based on class imbalance handling

To address the dilemma of insufficient recognition of the minority class caused by class imbalance in predicting depression tendency among college students, and systematically compare the predictive performance of 7 oversampling, undersampling, and hybrid sampling methods combined with 12 machine learning models on a class-imbalanced dataset of music APP listening habits and depression tendency in them, providing a reference for method selection in handling class-imbalanced data.

This study adopted a cross-sectional study design. Data were collected from questionnaire surveys of college students from 29 provinces, autonomous regions, and municipalities in China in 2023, with college students primarily being undergraduate and postgraduate students aged 18 to 24 years. Univariate analysis was used to screen 10 features significantly associated with depression tendency in the participants. Three oversampling methods including synthetic minority over-sampling technique (SMOTE), adaptive synthetic sampling approach (ADASYN), and synthetic minority over-sampling technique for regression with Gaussian noise (SMOGN), 2 undersampling methods including edited nearest neighbors (ENN) and Tomek Links, and 2 hybrid sampling methods including SMOTEENN and ADASYN-Tomek were applied to process the raw data. The raw data and processed data were respectively used to construct prediction models with 12 classification algorithms, including logistic regression (LR), support vector machine (SVM), random forest (RF), decision tree (DT), Light Gradient Boosting Machine (LightGBM), K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGBoost), Lasso logistic regression (LASSO), ridge regression (Ridge), elastic net (ENet), multilayer perceptron (MLP), and stacking ensemble model (Stacking). Area under the receiver operating characteristic curve (AUC), Recall, F1-score, and balanced accuracy were selected as evaluation metrics to assess model performance. The optimal model was screened by comparing various combinations of class-imbalance handling techniques and machine learning strategies. SHapley Additive exPlanations (SHAP) analysis was applied for interpretability analysis for the optimal model.

Results Univariate analysis identified 10 features significantly associated with depression tendency in college students: gender, grade, major category, preference for Children’s songs, approximate length since music listening habits, preference for Chinese style songs, usual music styles, daily listening duration, period of listening during a day, and frequency of comments on music APP. The comparison of 84 combined models with the 12 models based on raw data showed that hybrid sampling outperformed single sampling methods among the 7 resampling strategies; among the 12 classification algorithms, the stacking ensemble model demonstrated better comprehensive performance than individual classifiers. The combination of SMOTEENN hybrid sampling with Stacking achieved the best performance (AUC=0.95, F1=0.89, Recall=0.88, balanced accuracy=0.89). SHAP analysis showed that “usual listening style” contributed the highest in predicting depression tendency among college students.

Based on the comparison of 84 combination models, the SMOTEENN+stacking ensemble model can accurately predict depression tendency in college students. The comparative analysis provides a methodological basis and practical reference for method selection in class-imbalanced data scenarios and for early identification of mental health risk based on music listening behavior data.