Predictive Policing Using Deep Learning: A Community Policing Practical Case Study

There is relentless effort in combating the issue of crime in South Africa and many parts of the world. This challenge is heightened in under-resourced settings, where there is limited knowledge support, thus resulting in increasing negative perceptions of public safety. This work presents a predictive policing model as an addition to a burglar alarm system deployed in a community policing project to improve crime prevention performance. The proposed model uses feature-oriented data fusion method based on a deep learning crime prediction mechanism. Feed-Forward Neural Network (FFNN) and Recurrent Neural Network (RNN) models are employed to predict the amount of calls made to police stations on a monthly basis. Device installation and census data are used in the feature selection process to predict monthly calls to a police station. Coefficient of correlation function is used to isolate the relevant features for the analysis. To provide a viable way of achieving crime reduction targets, the models are implemented and tested on a real-life community policing network system called MeMeZa, which is currently deployed in low-income areas of South Africa. Furthermore, the model is evaluated using coefficient of determination function and the accuracy of the predictions assessed using an independent dataset that was not used in the models' development. The proposed solution falls under the Machine Learning and AI applications in networks paradigm, and promises to promote smart policing in under-resourced settings.