Using Machine Learning to construct an Internal Classification of Male Prison Inmates

• In Event: In God We Trust. All Others Must Bring Data

Abstract

The current status of Internal Classification (IC) in USA state prisons is first briefly reviewed to give a context for this project. The distinct goals of Internal and External Prison classifications are clarified. We then identify key design flaws that have undermined several prior IC systems and design goals for a new IC are specified. The sample consisted of 14, 049 Michigan state prisoners. Input classification factors were explicitly theory-guided to strengthen the explanatory, treatment and responsivity goals of the new IC. The pattern identification phase used unsupervised learning procedures (Bagged k-means and classic k-means). Multiple cross-validation approaches were used to identify the most reliable classification models for a new IC. The full classification hierarchy was examined to the most appropriate classification levels (K = 5 and 8) with cross validation at each level. Two classifiers (Random Forests and Support Vector machines) were then trained to automatically assign unknown (new) cases to their correct final IC category. The internal structure, homogeneity levels and meanings of core prototypes and relations between them are presented. These prototypes were then compared with the prior prison IC literature to identify commonalities across this field.

Authors

• Tim Brennan, Northpointe Institute

• William Dieterich, Equivant

• Christina Mendoza, Equivant