Predicting Recidivism With Machine Learning: An Analysis of Risk Factors and Proposal of Preventions

Abstract

Despite efforts to support the re-entry of prisoners into society, a significant proportion of released offenders eventually return to crime. To identify areas for improvement within correctional facilities, researchers are directing their focus towards recidivism, the tendency of an offender to recommit a crime. Although past studies have identified factors that are correlated to recidivism, there is still uncertainty about the most significant combinations of factors that drive it. Due to the complexity of this issue, the goal of our project is to create a machine-learning model to predict whether an individual will relapse into crime. Such a model will help experts study the effectiveness of specific forms of punishment and develop personalized correctional programs to target individuals based on their recidivism risk factors. We applied the Decision Tree, Random Forest, and Gradient Boosted Decision Tree algorithms to classify a prisoner as likely to recidivate or not, tuning the hyperparameters to optimize accuracy. To evaluate our hypotheses, we analyzed the top nodes of our trees and confirmed several of our initial predictions. Furthermore, we found that an individual’s relative placement in their community, such as the percentage of individuals in their community with lower or equal education levels, was a significant predictor of recidivism. The results of this research may help law enforcement officers make more informed decisions about how to allocate their resources based on predictions of recidivism.