Predicting Criminal Recidivism with Support Vector Machine

Abstract

Predicting criminal recidivism effectively is of major interest in criminology. In this paper, we study the ability of the support vector machines (SVM) to predict the probability of reincarceration. As a semi parametric approach, the SVM minimizes structural risk whereas nonparametric models, such as neural networks, minimize empirical risk. Furthermore, the SVM differs significantly from existing parametric models, such as logistic regression, in prediction of criminal recidivism. Due to the relatively new application of the SVM in predicting criminal recidivism in the field of criminology, a general framework is presented for how the SVM may become a supplemental or alternative method for recidivism prediction. Comparisons among logistic regression, neural networks, and the SVM are made with empirical testing results on a well-known recidivism data set. A combined prediction utilizing all three methods provides the most flexibility and accuracy in decision-making.