Predictive Policing—Are Ensemble Methods More Accurate Than Regression Methods?

Abstract

AbstractThe ability to predict crime before it takes place has acquired significant currency over the past decade. Predictive policing aims to achieve this ability using mathematical, predictive analytics in law enforcement to identify potential criminal activity. As machine learning models make inroads in predictive policing, this research aims to compare two widely used predictive policing methods with the concerted goal of predicting the time and locations where crime is most likely to occur—ensemble methods and individual regression. Leveraging historical datasets for two cities in the USA—San Francisco, CA, and Chicago, IL, we developed models for two distinct predictive policing techniques and compared their accuracy using a set of four metrics. Our analysis reveals the superiority of ensemble methods in their ability to accurately anticipate the occurrence or lack of crimes in each region, which is reflected by their significantly higher average true rate. Furthermore, the research also identifies crime distribution and spatiotemporal effects as potential reasons behind the disparity in accuracy levels.KeywordsMachine learningEnsemble methodsRegressionPredictive policing