Using Deep Learning to Forecast Spatiotemporal Crime Patterns

The distributional patterns of crime occurrences are closely related to their spatial, temporal, and environmental contexts. It has been a hot topic for researchers and crime analysts to discover such complex relationships in order to forecast crime, both spatially and temporally. Many factors play a role in the occurrences of crimes. Conventional crime forecasting research has primarily relied on historical crime records and socioeconomic data, while ignoring the rich social media and other environmental context data. The large volume of data requires a more appropriate forecasting framework with the ability to take in massive multimodal data and possibly achieve better predictive performance. In this dissertation research, the applicability of deep learning was extended to the field of crime forecasting, where researchers have done little work applying this powerful and high-performance technique. The proposed deep neural networks models are based on multimodal data sets and have shown superior predictive power than other baseline models. These deep learning models are especially designed to properly fuse the diverse data sets, which have various types, scales, and formats. In addition, the proposed deep neural networks models are capable of capturing both spatial and temporal dependencies of several input data features, which is crucial for crime forecasting research. In order for the deep learning architectures to fully take advantage of the multimodal data, advanced feature selection and text feature extraction methods are also employed. Hopefully, this research could assist crime analysts with obtaining better crime forecasting accuracy and thus would benefit the police patrolling allocation. Furthermore, it is possible that results of the proposed deep learning architectures could be extended to a range of other fields, where they have the properties of demonstrating spatiotemporal distributional patterns as well as containing large amounts of data