Predicting a few or many friends in schoolchildren: a machine learning approach

Machine learning algorithms were used to determine whether Peruvian 8-12-year-olds (N = 730) believe they have many or few friends based on sociodemographic and psychological variables. Theoretical variables were depression (PHQ-8), subjective well-being (SWB-3), sorrow, bullying, and birthplace. With stratified 10-fold cross-validation, the easyML R package examined six algorithms: Random Forest, XGBoost, Support Vector Machine, Neural Network, Logistic Regression, and Decision Tree. The cross-validation pipeline used SMOTE to correct class imbalance (25.6% few vs. 74.4% many friends). Logistic Regression had the greatest ROC-AUC (0.663 in cross-validation, 0.673 on the test set) and no overfitting (CV-test gap = -0.010). The greatest SHAP predictor was subjective well-being, followed by depression and birthplace. Youden's J index threshold optimization (optimal = 0.44) increased classification accuracy from 64.6% to 72.8%. Psychology predicts friendship perception in youngsters, with well-being outweighing clinical symptomatology.