Predicting Juvenile Crime Severity Using Machine Learning: A Study in Malaysian Rehabilitation Schools

Abstract

This study aims to classify juvenile crime severity using machine learning techniques based on data collected from 102 juvenile offenders across three rehabilitation schools under the Malaysian Department of Social Welfare. A self-administered survey captured demographic, psychosocial, and behavioral factors, while the target variable was categorized into three levels of crime severity. Three supervised classification models namely Logistic Regression, K-Nearest Neighbors (KNN), and Random Forest were applied to identify predictive patterns, and the Synthetic Minority Oversampling Technique (SMOTE) was employed to address class imbalance. Feature selection and ranking were conducted using Random Forest feature importance, with statistical validation provided through multinomial logistic regression p-values. The results indicate that Random Forest achieved the highest predictive performance, attaining an accuracy of 58.1% before SMOTE and 51.6% after SMOTE, outperforming Logistic Regression (45.2%) and KNN (41.9%). When restricted to the six most important predictors which are Family Background, Emotional Health, Gender, Experience of Premarital Sex, Age Group, and Education Level, the Random Forest model maintained its peak accuracy of 58.1%, indicating that a reduced-feature model can preserve predictive performance while improving interpretability. Within this subset, Family Background and Emotional Health were identified as the most influential features by the Random Forest model. In parallel, multinomial logistic regression analysis identified Gender, Experience of Premarital Sex, and Age Group as statistically significant predictors of crime severity (p < 0.05). Overall, the findings suggest that a combination of demographic, behavioral, and psychosocial factors plays a significant role in differentiating juvenile crime severity among Malaysian offenders. This study highlights the potential of AI-driven predictive modelling to support early risk identification, targeted rehabilitation planning, and evidence-based policy development within juvenile justice management.