Algorithmic determination of recidivism outcomes for improved risk prediction

Abstract

Under European Union’s Artificial Intelligence (AI) Act, high-risk AI systems such as recidivism risk-assessment tools must perform accurately with minimal residual risk. However, such tools often rely on outdated training data due to the complex and time-consuming nature of identifying post-release outcomes. This not only prevents timely evaluations, but also undermines regulatory compliance. A notable example is RisCanvi, a recidivism risk-assessment instrument that has been in use across prisons in Catalunya since 2009. In this thesis, we address these limitations for RisCanvi by translating legal rules into a certifiably accurate algorithm (validated with domain experts) for determining recidivism. Leveraging this algorithm, we construct a high-quality dataset linking post-release outcomes with RisCanvi evaluations for 17.8K inmates released in Catalunya between 2010 and 2019. This dataset significantly improves the prediction of both general and violent recidivism relative to prior versions of RisCanvi, while enabling a comprehensive evaluation of model configurations. Drawing on this analysis, we offer empirically grounded recommendations on training data size, class re-weighting strategies, and model selection for optimizing predictive performance. Through an extensive ablation study, we identify features unrelated to criminal behavior and/or not within the inmates’ control that are redundant for predicting recidivism risk. Importantly, we demonstrate that imposing monotonicity shape constraints on RisCanvi’s features preserves predictive performance while ensuring that rehabilitative progress lowers predicted risk. Finally, we show that this constrained model satisfies relaxed algorithmic fairness constraints and produces subgroup-invariant risk scores, making it a suitable tool for decision support in high-stakes settings.