An Interpretable Dual-Model Framework: Integrating Gradient Boosting and Random Forest to Decode Key Adolescent Stressors in Educational Contexts

Abstract

Student stress prediction requires capturing complex interactions among psychological, physiological, and environmental factors. This study develops an interpretable dual-model framework integrating Gradient Boosting Machine (GBM) and Random Forest (RF) to identify key stressors and their underlying mechanisms in adolescents. Using data from 1,000 adolescents, we conducted hyperparameter optimization (GBM: learning rate = 0.08, max depth = 4; RF: mtry = 6, max depth = 5) and multi-modal validation (Spearman correlations, SHapley Additive exPlanations [SHAP] analysis, and feature importance rankings). Key results reveal: (1) Both models achieved high predictive accuracy (R² > 0.80, MAE < 0.15); (2) Self-esteem emerged as the dominant stress predictor (ΔR² ≈ 0.13, acting as a stress buffer), followed by academic performance (ΔR² ≈ 0.11); (3) SHAP visualizations uncovered nonlinear threshold effects (e.g., academic performance) and anxiety-mediated pathways; (4) RF demonstrated superior noise robustness (MAE = 0.135 vs. GBM’s 0.146), while GBM better captured linear relationships in physiological variables. This framework enables targeted stress intervention strategies through feature importance rankings, significantly optimizing resource allocation in educational mental health programs.