Machine learning algorithms for crime prevention and predictive policing

Recent developments within the field of Machine Learning have given rise to the possibility of deploying these algorithms within a live policing environment. This thesis, motivated by the needs of Dyfed-Powys Police, focuses on developing a series of predictive tools that can be used directly within a live setting in order to improve efficiency across the force. With an area of coverage that spans four socioeconomically diverse yet sparsely populated counties, Dyfed-Powys Police face a unique set of challenges in managing an increasingly limited set of resources such that offenders can be properly managed. The issue of personnel management is first addressed in the construction of a recommender system, which investigates the use of clustering techniques to exploit a stable pattern in the times at which crimes occur in various locations across the region. This is then followed with the development of a Recurrent Neural Network, which aims to predict the time to next offence within a particular narrowly-defined partition of the area. By developing a series of tools that make use of existing data to predict which offenders within their database are most likely to reoffend, we aim to assist Dyfed-Powys in monitoring and preventing recidivism across the area. Firstly, we investigate the use of Random Forests and XGBoost algorithms, as well as Feedforward Neural Networks to predict an offender's likelihood of reoffence from a series of diverse factors. Secondly, we develop the aforementioned Random Forests algorithm into a survival model that aims to predict an offender's time to reoffence. Lastly, we develop a stacked model, which uses publicly available data to construct an Area Classifcation score for use as a factor within the original reoffence classification model. Insightful results are obtained, indicating a clear case for the use of many of these techniques in a live setting