Criminal recidivism. Towards reliable and transparent predictive models

Abstract

The prediction of criminal recidivism through machine learning (ML) models raises significant ethical, legal, and methodological challenges. This article promotes for a transparency and explainability-oriented approach by comparing three predictive models – logistic regression, random forest, and neural networks – applied to the COMPAS dataset of criminal history data, released by ProPublica a non-profit journalism organization in USA. To assess the coherence and readability of algorithmic decisions interpretability techniques such as SHAP values are employed. The analysis also considers the implications of adjusting the decision threshold to increase false positives for supportive – rather than punitive – purposes, emphasizing the greater ethical and social acceptability of such a strategy. The discussion is complemented by an overview of the regulatory developments in Italy and the European Union regarding the use of predictive technologies in the criminal justice system.