AI and prediction of criminal justice risk: revisiting the COMPAS dataset, experiments, and the problem of algorithmic bias and fairness

The aim of this study is to evaluate the accuracy of machine learning techniques in addressing the problem of bias in criminal assessment prediction. Previous studies have shown that commercial software for criminal risk assessment produces biased predictions, no better than assessment by randomly selected people without criminal justice expertise (Dressel and Farid in Sci Adv 4(1):eaao5580, 2018. https://doi.org/10.1126/sciadv.aao5580 ). In this research, we reproduce the results of Dressel and Farid (Sci Adv 4(1):eaao5580, 2018. https://doi.org/10.1126/sciadv.aao5580 ), testing Logistic Regression (LR) and Support Vector Machines (SVM) on the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) dataset. In addition, we implement an eXtreme Gradient Boosting (XGB) classifier and study the effect of hyper-parameter optimization and correlation-remover as debiasing techniques. We find that XGB performs better than Logistic Regression, and that debiasing techniques modestly improve accuracy of prediction. We conclude that bias mitigation techniques can be helpful up to a point but note that embedded bias in the data can persist in AI risk assessment tools, which can have profound social and ethical implications for individuals and for society.