Crime Hot Spot Forecasting: A Recurrent Model with Spatial and Temporal Information

Crime is a major social problem in the United States, threatening public safety and disrupting the economy. Understanding patterns in criminal activity allows for the prediction of future high-risk crime 'hot spots' and enables police precincts to more effectively allocate officers to prevent or respond to incidents. With the ever-increasing ability of states and organizations to collect and store detailed data tracking crime occurrence, a significant amount of data with spatial and temporal information has been collected. How to use the benefit of massive spatial-temporal information to precisely predict the regional crime rates becomes necessary. The recurrent neural network model has been widely proven effective for detecting the temporal patterns in a time series. In this study, we propose the Spatio-Temporal neural network (STNN) to precisely forecast crime hot spots with embedding spatial information. We evaluate the model using call-for-service data provided by the Portland, Oregon Police Bureau (PPB) for a 5-year period from March 2012 through the end of December 2016. We show that our STNN model outperforms a number of classical machine learning approaches and some alternative neural network architectures.