A fairness-focused approach to recidivism prediction

Implications for accuracy, trust, and equity

Bibliographic Data

ID20397143
AuthorsMichael Mayowa Farayola (0009-0002-3495-4155, Dublin City University, corresponding author), Irina Tal (0000-0001-9656-668X, Dublin City University), Takfarinas Saber (0000-0003-2958-7979, Ollscoil na Gaillimhe – University of Galway), Robert Connolly (0000-0003-3196-2889, Dublin City University), Regina Connolly, Malika Bendechache (0000-0003-0069-1860, Ollscoil na Gaillimhe – University of Galway)
Year2026
Volume41
Issue4
Pages2783-2801
Publication date2026-04-01
Open AccessYes
TypeARTICLE
VenueAI & Society (JOURNAL)
Journal identifiersISSN: 0951-5666 • E-ISSN: 1435-5655
PublisherSpringer Science and Business Media LLC (PUBLISHER)
DOI10.1007/s00146-025-02452-1
OpenAlexW4412358206
LanguageEN
Citations received2
References cited44

Fairness in AI has emerged as a critical priority, particularly in high-stakes domains like criminal justice, where algorithmic decisions profoundly affect individuals and demographic groups. This study addresses persistent challenges of bias in recidivism prediction by evaluating the integrated application of fairness-enhancing techniques across the AI pipeline. Specifically, we analyze the integrated use of Reweighing, Adversarial Learning, Disparate Impact Remover, Exponentiated Gradient Reduction, Reject Option Classification, and Equalized Odds optimization. These methods are assessed using key fairness metrics, Statistical Parity Difference (SPD), Disparate Impact (DI), Equal Opportunity Difference (EOD), and Predictive Equality Difference (PED), on two real-world recidivism datasets: COMPAS, a well-known publicly available dataset, and RisCanvi, a curated English dataset derived from a publicly available Spanish dataset. Our findings reveal trade-offs between fairness improvements and predictive accuracy, with fairness performance varying across integrated mitigation strategies. While individual techniques often fall short of effectively reducing bias, their integration across pre-processing, in-processing, and post-processing stages demonstrates a more comprehensive and robust capacity to address fairness concerns. In some instances, fairness mitigation strategies exacerbate disparities instead of reducing them, highlighting the complexity of achieving equitable AI outcomes. Some methods struggle with specific datasets, leading to worsened SPD, DI, EOD, or PED values, potentially reinforcing biases rather than eliminating them. Using multi-objective and bi-objective optimization frameworks, we identify optimal configurations that balance fairness and accuracy, contributing to a deeper understanding of fairness–accuracy trade-offs in AI systems. This study makes a novel contribution by systematically evaluating multi-phase fairness integration in recidivism prediction, bridging critical gaps in the literature. However, our results also underscore the risks of ineffective or counterproductive fairness interventions when not carefully tailored. Our findings highlight the necessity of tailoring bias mitigation strategies to specific datasets and contexts, providing actionable insights for developing equitable, diverse, and inclusive AI systems that are fair and reliable in the criminal justice system

Actuarial science · Business · Criminology · Political science · Recidivism · Computer Science · Ethics and Social Impacts of AI · Ethics in Business and Education · Law · Psychology · Psychology of Moral and Emotional Judgment

Tools

Open DOIOpen Access