Predicting Recidivism Risk Using Explainable Machine Learning Models: A Case Study from Eastern European Correctional Data

Povzetek

Recidivism prediction is a critical component of modern criminal justice systems, especially in efforts to design more effective rehabilitation and risk assessment strategies. This study applies explainable machine learning (ML) techniques to analyze recidivism risk among formerly incarcerated individuals using a real-world correctional dataset from an Eastern European country. We compare the predictive performance of several models, including logistic regression, random forest, and XGBoost, with a particular focus on model interpretability using SHAP (SHapley Additive exPlanations) values. The results show that ML models significantly outperform traditional statistical approaches in predictive accuracy, with XGBoost achieving the highest performance (AUC = 0.83). Moreover, explainability analysis reveals that factors such as age at release, prior offense type, educational attainment, and employment status are key predictors of reoffending. The study highlights the potential of interpretable ML tools to support data-driven decision-making in corrections and parole systems, ensuring both predictive power and transparency.