PSC Conference Room, McNeil 574
Title: Optimized Scoring Systems in Criminal Justice
Abstract: Scoring systems are linear classification models that let users make quick predictions by adding, subtracting and multiplying a few small numbers. Starting with the work of Burgess in 1928, these models have been used to support high-stakes decisions in criminal justice, from setting bail, to sentencing, to determining release on parole. In spite of extensive use, however, scoring systems are still built using ad hoc approaches that hinder predictive performance and fail to address other constraints required for effective deployment.
In this talk, I present two new methods to learn optimized scoring systems from data, first for decision-making (SLIM - Supersparse Linear Integer Models), and then for risk assessment (RiskSLIM - Risk-calibrated Supersparse Linear Integer Models). Unlike traditional methods for predictive modeling, both methods require solving hard computational problems to optimize and constrain exact measures of performance and form (e.g., the number of mistakes and predictors).
I describe how we can solve these optimization problems, and thereby obtain scoring systems that are fully optimized for performance, sparsity, integer coefficients, and other operational constraints. I demonstrate the benefits of this approach by building optimized scoring systems for recidivism prediction - showing that these simple models perform just as well as powerful machine learning models, but are far easier to use, understand, and customize. I end by discussing extensions of this work to improve the transparency, accountability, and fairness of predictive models in criminal justice.
Bio: Berk Ustun is a PhD candidate in Electrical Engineering and Computer Science at MIT. His research interests broadly include optimization, statistics, and machine learning. His work focuses on developing new methods for data-driven decision making in criminology and healthcare. To date, he has worked with the Harvard School of Public Health, the US Army STARRS, and multiple departments at the Massachusetts General Hospital. Prior to starting his PhD, Berk completed a MS in Computation for Design and Optimization at MIT, and BS degrees in Economics and Operations Research at UC Berkeley.