Predicting juvenile offending: a comparison of data mining methods.

Area of Science:

• Forensic Psychology
• Machine Learning in Criminology
• Adolescent Offending Behavior

Background:

• Juvenile offending is a significant societal concern.
• Accurate prediction of initial offending is crucial for intervention.
• Data mining techniques offer potential for improved predictive accuracy.

Purpose of the Study:

• To compare the efficacy of logistic regression and predictive data mining techniques (decision trees, artificial neural networks, support vector machines) in discriminating initial juvenile offending.
• To identify consistent risk factors for juvenile offending across different predictive models.
• To validate predictive models in independent Asian samples.

Main Methods:

• Logistic regression, decision trees (DTs), artificial neural networks (ANNs), and support vector machines (SVMs) were employed.
• Models were trained and validated on a large Asian sample, with independent testing.
• Receiver operating characteristic (ROC) analyses were used to evaluate classifier performance.

Main Results:

• All classifiers, including logistic regression, DTs, ANNs, and SVMs, achieved accuracy rates of 95% and above.
• Proactive aggression and teacher-rated conflict consistently emerged as significant risk factors across models and datasets.
• Other identified risk factors included reactive aggression, narcissistic traits, male gender, and non-intact family status, though with less consistency.

Conclusions:

• Predictive data mining and logistic regression are highly effective in identifying adolescents at risk for initial offending.
• Proactive aggression and teacher-rated conflict are robust predictors of juvenile offending.
• The findings support the use of advanced computational methods for risk assessment in juvenile justice systems.