Social Determinants of Recidivism: A Machine Learning Solution.

In criminal justice analytics, the widely-studied problem of recidivism prediction (forecasting re-offenses after release or parole) is fraught with ethical missteps. In particular, Machine Learning (ML) models rely on historical patterns of behavior to predict future outcomes, engendering a vicious feedback loop of recidivism and incarceration. This paper repurposes ML to instead identify social factors that can serve as levers to prevent recidivism. Our contributions are along three dimensions. (1) Recidivism models typically agglomerate individuals into one dataset, but we invoke unsupervised learning to extract homogeneous subgroups with similar features. (2) We then apply subgroup-level supervised learning to determine factors correlated to recidivism. (3) We therefore shift the focus from predicting which individuals will re-offend to identifying broader underlying factors that explain recidivism, with the goal of informing preventative policy intervention. We demonstrate that this approach can guide the ethical application of ML using real-world data.