Predicting Recidivism with Machine Learning Algorithms: A Comparison of Models

• In Event: Administrative Data and Predictions

Abstract

We present results from ongoing research on predicting recidivism with machine learning algorithms. The algorithms include Lasso regression, Random Forest classification (RF) and Generalized Boosted Regression (GBM). Our initial dataset consisted of a large reentry sample from a Midwestern state. The data included information on criminal history gathered from the state's management information systems and other information collected with the COMPAS assessment software. The two outcomes predicted by the models were arrest within three years after release for a violent felony offense and for a non-violent felony offense. Our findings indicate that both the RF and GBM algorithms yield models with comparable predictive validity and that Lasso regression, although not quite as discriminating, yield models that are parsimonious and straightforward to interpret by practitioners. We will discuss issues that arise when comparing models on predictive validity, such as what measures should be given the greatest weight and to what extent the process of model tuning affects the interpretation of the results. When the meeting occurs, we expect to have added to our findings by including other machine algorithms in the model comparisons and by applying the algorithms to a probation sample.

Authors

• William L. Oliver, Northpointe, Inc.

• William Dieterich, Northpointe, Inc.

• Tim Brennan, Northpointe, Inc.