Singular race models: addressing bias and accuracy in predicting prisoner recidivism

Abstract

As machine learning based predictive systems pervade many aspects of our lives, an inherent bias and unfairness surface from time to time in the form of mispredictions in various domains. Recidivism, the tendency of offenders to reoffend after release from prison on parole, is one such domain where one race-based sub-population has been found to be treated more harshly than others. Current practices have focused on eliminating race information from datasets to reduce the predictive bias. In contrast to this, we built Singular Race Models, a novel approach of segmenting the dataset based on race, to train and test single race-based models to increase prediction accuracy and reduce racially inspired bias by considering only one race at a time. We created Singular Race Models for four different crime categories and compared these with base models created using all crimes and all races. This modeling choice helped us increase accuracy and analyze race related discrimination. A three-layered artificial neural network was utilized to do the heavy weight-lifting of recidivism prediction. With the help of several suitable metrics, in this paper, we demonstrate the increase in predictive accuracy of these Singular Race Models in various crime categories and analyze the causes and the secondary effect on bias.