Improving Recidivism Forecasting With a Relaxed Naïve Bayes Classifier

Abstract

Correctional authorities require accurate, unbiased, and interpretable tools to predict individuals’ chances of recidivating if released into the community. However, existing prediction models have serious limitations meeting these requirements. We overcome these limitations by applying an established medical diagnostic approach: a relaxed naïve Bayes classifier. Using logistic regression in the form of a naïve Bayes classifier, we estimate the weights of observed features of offenders on recidivism. We apply these weights in a relaxed naïve Bayes classifier to predict the probability of recidivism. Results show that acquired features are stronger predictors of recidivism than innate features. Relaxed naïve Bayes classifier produces far less racial disparity than most alternatives. Critically, it is easier for users to interpret than its alternatives.