Abstract:
|
I will present evidence that black box proprietary models used currently in the criminal justice system can be replaced with transparent machine learning models without losing accuracy. This begs the question of why we allow companies to produce black box models and sell them to the criminal justice system. However, if we are to replace all black box models with interpretable ones, we require a toolbox of modern statistical techniques for transparent machine learning. In this talk, I will provide some of these optimization-based tools and their application to recidivism risk prediction.
|