Individual Submission Summary
Share...

Direct link:

The Pros and Cons of Machine-Learning Model Stacking for Predicting Criminal Justice Outcomes

Thu, Nov 17, 11:00am to 12:20pm, L404, Lobby Level

Abstract

Our team's primary approach was to use xGBoosted decision trees. Despite our lack of knowledge regarding how best to tune such a model, it performed reasonably well in predicting Year 1 and Year 2 recidivism. We shifted gears for our Year 3 predicting and did model stacking. This approach averages the predictions made by competing models. We generated two competing xGBoosted models and two neural network models. Our final Year 3 probabilities were based on all four models and performed very well. In general, this type of stacking has performance advantages over a single model approach when model precision is paramount. A disadvantage of such an approach is creating a model that is too complex to be used in a real-world setting. Even a single xGBoosted model would be challenging to use in a criminal justice setting. We provide a brief overview of other model specification details and implications of this modeling approach.

Authors