In this study, we propose advancements in criminal justice analytics along three dimensions. First, for the long-standing problem of recidivism risk assessment, we shift the focus from predicting the likelihood of recidivism to identifying its underlying determinants within distinct subgroups. Second, to achieve this, we introduce a machine learning pipeline that combines unsupervised and supervised techniques to identify homogeneous clusters of individuals and find statistically significant determinants of recidivism within each cluster. We demonstrate useful heuristics to address key challenges in this pipeline related to parameter selection and data processing. Third, we use these results to compare outcomes across subgroups, enabling a more nuanced understanding of the root factors that lead to differences in recidivism. Overall, this approach aims to explore new ways of addressing long-standing criminal justice challenges, providing a reliable framework for informed policy intervention.