Crime event prediction with dynamic features

Abstract

Nowadays, Location-Based Social Networks (LBSN) collect a vast range of information which can help us to understand the regional dynamics (i.e. human mobility) across an entire city. LBSN provides unprecedented opportunities to tackle various social problems. In this work, we explore dynamic features derived from Foursquare check-in data in short-term crime event prediction with fine spatio-temporal granularity. While crime event prediction has been investigated widely due to its social importance, its success rate is far from satisfactory. The existing studies rely on relatively static features such as regional characteristics, demographic information and the topics obtained from tweets but very few studies focus on exploring human mobility through social media. In this study, we identify a number of dynamic features based on the research findings in Criminology, and report their correlations with different types of crime events. In particular, we observe that some types of crime events are more highly correlated to the dynamic features, e.g., Theft, Drug Offence, Fraud, Unlawful Entry and Assault than others e.g. Traffic Related Offence. A key challenge of the research is that the dynamic information is very sparse compared to the relatively static information. To address this issue, we develop a matrix factorization based approach to estimate the missing dynamic features across the city. Interestingly, the estimated dynamic features still maintain the correlation with crime occurrence across different types. We evaluate the proposed methods in different time intervals. The results verify that the crime prediction performance can be significantly improved with the inclusion of dynamic features across different types of crime events.