Prison data analysis finds choosing lowest reoffending risk scores preserves overall accuracy
Whether an individual reoffends after their release from prison or is successfully rehabilitated depends on several factors. When assessing the risk of recidivism, all of these are taken into account—both by human experts and by machine learning models. If several equally reliable computational models produce different risk ratings, the lowest value should be used for the individual concerned,…
A new analysis of data from Catalonia's prison system has found that using the lowest risk score from multiple equally reliable machine learning models preserves overall accuracy. The study, conducted by Ashwin Singh of TU Wien and Carlos Castillo of ICREA/Universitat Pompeu Fabra, investigated the RisCanvi system, which has been used in Catalonia's prisons since 2009.
When multiple models produce slightly different risk ratings for the same individual, the team recommends selecting the lowest score to minimize ambiguity and ensure fairness. The researchers examined the RisCanvi system, which considers various factors including criminal history and education level, to predict recidivism risk. They developed interpretable linear models that improved prediction performance while reducing disparities in error rates across demographic groups.
Crucially, selecting the lowest risk score from multiple models did not compromise overall predictive performance. The study highlights the importance of model stability in high-risk applications, as well as the need for human oversight in decision-making processes.
Written by urgent.news from Phys.org's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
