Predicting Recidivism Through Machine Learning

The current study attempts to explore more precise solutions for the prediction problem in the U.S. criminal justice system. Using data obtained from the Bureau of Justice Statistics (BJS), Recidivism of Prisoners Released in 1994, this project investigates a number of statistical models in terms of their classification performances. Specifically, after a careful feature selection process, the conventional logistic regression model is compared to several machine learning models, including random forests, support vector machines, XGBoost, neural networks, and Search algorithm. According to the results, XGBoost and neural networks outperformed all other models in the comparison. Implications for the U.S. criminal justice system are discussed.