Toward precision prevention: machine learning-identified risk and protective factors for distinct bullying...

Bullying victimization among Chinese adolescents manifests in distinct types (verbal, physical, relational, cyberbullying) with unique adverse outcomes and influencing factors, yet predictive models for specific types remain scarce. Collapsing these distinct victimization types into a single outcome may obscure critical differences essential for tailored prevention. Therefore, this study applied machine learning to predict distinct victimization types and reveal type-specific risk and protective factor patterns, supplying an evidence base for targeted intervention.

Chinese adolescents (n = 1,981, aged 11-18 years) completed measures of bullying victimization, depression, anxiety, friendship quality, emotion regulation, social support, school climate, trait anger, and moral disengagement bullying behavior. Multiple machine learning algorithms were trained using 5-fold cross-validation, with sensitivity analyses conducted without oversampling. For each victimization type, the hyperparameter configuration yielding the highest F2-score was selected to retrain the model on the full training set and subsequently applied to the test set. SHapley Additive exPlanations (SHAP) analysis was performed on the F2-optimal model using test set data to identify key predictors, and feature importance rankings were compared between the main and sensitivity analyses.

Random Forest achieved optimal performance for overall (F2 = 0.67) and physical (F2 = 0.52) victimization, Bagging for verbal (F2 = 0.65), Logistic Regression for relational (F2 = 0.57), and Naïve Bayes for cyberbullying (F2 = 0.51). Depression, anxiety, angry reaction and moral disengagement were common risk factors, while student-student was relationship a common protective factor, while angry temperament, gender, conflict and betrayal, and teacher-related factors, among others, showed distinct patterns across victimization type. Sensitivity analyses confirmed these patterns: the top five predictors were consistent across verbal, physical, cyber, and overall victimization, and four of the top five were also consistent for relational victimization.

This study demonstrates that machine learning enables prediction of distinct bullying types and identification of their unique risk and protective factor patterns. These findings suggest that prevention strategies should move beyond one-size-fits-all approaches toward tailored interventions addressing the specific combination of risk factors of each victimization type.