Machine Learning Approaches as a Tool for Effective Offender Risk Prediction

• Date: 01 August 2013
• DOI: http://doi.org/10.1111/1745-9133.12060
• Author: William Rhodes
• Published date: 01 August 2013

EDITORIAL INTRODUCTION

FORECASTING CRIMINAL BEHAVIOR

Machine Learning Approaches as a Tool for

Effective Offender Risk Prediction

William Rhodes

Abt Associates

The prediction of criminal behavior plays an instrumental role in criminal justice admin-

istration (Gottfredson and Moriarty, 2006). Social workers provide high-risk youth with

mentoring, police intensify patrol activity in high-risk neighborhoods, prison administrators

segregate high-risk offenders from low-risk ones, and community corrections administra-

tors concentrate controlling and correctional resources on supervisees at an elevated risk

of recidivism. Risk assessment is an essential ingredient of evidence-based criminal justice

administration, and good risk assessment is better than bad risk assessment.

Berk and Bleich (2013, this issue) advocatefor wider use of machine learning approaches

for deriving predictions. This advocacy appears as a gentle introduction intended to convince

readers that machine learning works, to motivate a deeper reading into a technical literature

for why it works (Berk, 2012), and eventually to induce widespread application. Their

argument has two principal components.

Berk and Bleich (2013) assert that predication is difficult and uncertain using conven-

tional regression-based methods. They reference a complex decision boundary, by which

they mean that good prediction can depend on many variables, which might require

transformations, interactions, and nonlinear manipulation of data. They are skeptical of

regression-based adaptations for identifying complex decision boundaries, and even if in

theory regression-based procedures could be hammered into a suitable form, the training

and experience of most criminal justice researchers provide inadequate statistical carpentry.

Do not take a chance on conventional regression-based procedures, they exhort; turn to

machine learning algorithms that can detect complex patterns.

The second component of Berk and Bleich’s (2013) argument is that regression-based

approaches use an inadequate loss function that fails to weight some outcomes as more

serious than others (Berk, 2011). For example, predictions placing greater weight on future

Direct correspondence to William Rhodes, Abt Associates, 55 Wheeler Street, Cambridge, MA 02138 (e-mail:

bill_rhodes@abtassoc.com).

DOI:10.1111/1745-9133.12060 C2013 American Society of Criminology 507

Criminology & Public Policy rVolume 12 rIssue 3