Evaluating the Efficacy of Stacked Ensemble Models in Predicting Failure to Appear

• In Event: Advances in Machine Learning and Crime

Abstract

There is a growing body of literature in criminology comparing the predictive accuracy of machine learning models to traditional statistical techniques. This literature finds mixed support for the efficacy of machine learning models compared to more common methods. Methods that combine individual models into a single model, however, have largely been ignored. First suggested by Wolpert (1992), stacked generalization models have the possibility to outperform an individual predictor by combining them into a larger model. Recent simulation research suggests an ensemble of logistic regression and random forests can produce a more resilient predictor than either one alone (Sibley, 2012). In predicting failure to appear in court, the current study evaluates the accuracy of several stacked ensemble models compared to the individual models from which they are derived. In doing so, this study adds stacked ensemble models to the growing body of literature on the efficacy of machine learning techniques for criminological applications.

Author

• Stephen Clipper, University of Alabama