Offense Severity Prediction Under Partial Knowledge: Trigger Factor Detection Using Machine Learning and Network Science Methods

Predictive policing in the new era requires a more accurate prediction of potential crime events. When data scientists use crime data to conduct analyses or prediction tasks, due to the particularity of crime data, the incompleteness of crime data has always been a big challenge. The purpose of this work is to find the most sensitive presumptive feature when only limited or delayed offense information is available. In this study, the authors create a framework that employs both machine learning (decision tree) and network science (eigenvector centrality) methods, to detect the trigger feature that has the greatest influence on the prediction of crime severity under only partial knowledge. The outcome of this work reveals the trigger features that can best improve the prediction results under different a priori knowledge contexts and provides a new evaluation indicator of crime hotspots for predictive policing.