Running summarizing tools on a new article

Abstract: The mental health challenges of college students have drawn extensive attention, and the accurate prediction of psychological resilience is the key to targeted mental health intervention. Traditional subjective scale assessment has technical defects such as low timeliness and poor predictive accuracy, while existing machine learning-based mental health prediction models suffer from single data dimension, low training efficiency and poor model interpretability. To address these technical problems in computer-aided mental health assessment, this study develops a machine learning algorithm-driven predictive model for college students' psychological resilience, and achieves three key computer technology advancements in the research: First, an optimized multi-source heterogeneous data preprocessing framework is proposed, which integrates LASSO regression feature dimensionality reduction and TF-IDF sentiment polarity analysis to realize efficient cleaning and feature extraction of mixed structured/unstructured psychological data, and the internal consistency of the dataset (Cronbach's α) is improved from 0.72 to 0.89. Second, the XGBoost ensemble learning algorithm is optimized by combining Bayesian optimization for hyperparameter tuning, which solves the low efficiency problem of traditional grid search; the model's R² score on the validation set is increased by 12.7%, and the training time is reduced to 35% of the grid search method. Third, the SHAP (SHapley Additive exPlanations) value interpretation method is introduced into the psychological resilience prediction model, which makes up for the defect of poor interpretability of complex ensemble learning models and quantifies the contribution of core features to prediction results. This study elaborates on the technical implementation details including data preprocessing procedures, algorithm parameter setting, model training strategies and performance evaluation metrics. Experimental results show that the proposed model achieves a predictive correlation coefficient of 0.84 and an AUC (Area Under ROC Curve) of 0.89 on the test set, which is significantly superior to traditional linear regression and mainstream machine learning algorithms (SVM, Random Forest) (p<0.01). Case studies further verify the technical robustness and practical applicability of the model in complex scenarios. The research results not only provide objective and quantifiable technical support for university mental health services, but also optimize the application framework of machine learning algorithms in the field of psychological assessment, and provide a feasible technical solution for the integration of multi-source heterogeneous data in mental health research.