A Machine Learning Model for Data-Driven Prediction of School Violence

Objectives This study aims to define relevant data attributes and develop an optimal predictive model to forecast the likelihood of school violence. Furthermore, it seeks to analyze the attributes contributing to school violence prediction and provide implications for future data-driven school violence prevention policy research. Methods Based on a review of prior studies, data attributes related to school violence prediction were defined. Using these defined attributes, data were collected from 1,274 elementary and secondary school students in Seoul. Attribute engineering techniques—such as handling missing values, feature selection, and transformation—were applied, resulting in four distinct datasets. Various machine learning models, including logistic regression, random forest, and gradient boosting, were employed. The AUC-ROC was used as the primary performance metric, and hyperparameters were optimized via grid search. Results The optimized models developed through this research demonstrated up to a 13.24% improvement in performance compared to baseline models, with the gradient boosting model achieving the best performance. SHAP analysis revealed that aggressiveness, impulsivity, parent-child relationships, and school-related conversations were key attributes contributing to the prediction of school violence. Conclusions The machine learning-based prediction model developed in this study enables quantitative estimation of the likelihood of perpetrating school violence. The findings also offer direction for early intervention and customized preventive education by considering key contributing attributes. This research provides a foundational basis for establishing and implementing data-driven school violence prevention policies.