Interpretable Machine Learning and Criminological Theories: Global Evidence on Bullying Perpetration and Victimization (2001–2014)

Abstract

While existing criminological theories offer valuable insights into the risk factors associated with bullying perpetration and victimization, further empirical assessments are needed—particularly across diverse temporal and cultural contexts. This study applies interpretable machine learning (IML), specifically random forest algorithms with feature importance measures, to explore the predictive relevance of key factors using four waves (2001–2014) of the Health Behaviour in School-Aged Children (HBSC) survey across approximately 40 countries. The findings reveal that antisocial lifestyle factors are the most salient predictors of bullying perpetration, whereas physical and psychological traits are more strongly associated with victimization. These patterns demonstrate notable consistency across both time and region, reinforcing the applicability of existing theoretical frameworks. By using the transparency of IML, this study not only evaluates core theoretical claims but also contributes to the development of targeted, evidence-based policies and interventions for bullying prevention in school settings.