Dissecting the Predictors of Cyber-Aggression Through an Explainable Machine Learning Model

The general aggression model (GAM) suggests that cyber-aggression stems from individual characteristics and situational contexts. Previous studies have focused on limited factors using linear models, leading to oversimplified predictions. This study used the light gradient boosting machine (LightGBM) to identify and rank the importance of various risk and protective factors in cyber-aggression. The SHAP (SHapley Additive exPlanations) technique estimated each variable's predictive effects, and two-dimensional partial dependence (PD) Plots examined interactions among predictors. Among 30 potential factors, the top five were attitudes toward violence, revenge motivation, anti-bullying attitudes, moral disengagement, and anger rumination. PD analysis showed significant interactions between protective factors (anti-bullying attitudes and moral reasoning) and risk factors (attitudes toward violence, revenge motivation, moral disengagement, and anger rumination). High scores on protective factors mitigated the impact of risk factors on cyber-aggression. These findings support and expand GAM, offering implications for reducing cyber-aggression among Chinese college students.