This submission to the National Institute of Justice’s (NIJ’s) 2021 Recidivism Forecasting Challenge describes the implementation of a deep learning model, specifically a neural net, to help predict recidivism and assist corrections officers in identifying high-risk individuals, thus helping to prevent recidivism.
Many current models also suffer from racial bias, so minimizing this bias is an essential element of the challenge. The model was trained on the data provided by the NIJ for the challenge. This dataset included approximately 2,600 persons on parole in Georgia through the years 2013 to 2015. The model was trained on 70 percent of information provided as “training” data for project submission. Despite its accuracy in generating predictions, the neural network model used is unable to provide information about which variables are more relevant than others in predicting recidivism. This is because the computations applied by each layer of the neural network take into account multiple variables from the prior layer of the neural network, causing the aforementioned abstraction from the provided input layers. The neural network created as an entry for this project was able to predict recidivism with a 67.73 percent accuracy, despite being a relatively simple model.
- Reducing Gun Violence through Integrated Forensic Evidence Collection, Analysis and Sharing
- Consideration of the probative value of single donor 15-plex STR profiles in UK populations and its presentation in UK courts
- Gas chromatography/mass spectrometry analysis of the six-ring regioisomeric dimethoxybenzyl-N-methylpiperazines (DMBMPs)