A deep learning framework for predicting burglaries based on multiple contextual factors

One of the most common types of crime, burglary often results in serious psychological trauma and has financial consequences. Predicting burglaries is a challenging task due to their high degree of randomness. In this study, we propose predicting burglaries based on various contextual factors and incorporating these factors in a unique deep learning framework – DeePrison (DEEp learning framework for Predicting buRglarIes baSed On multiple coNtextual factors). DeePrison is able to capture the dependencies between burglaries and their different contextual factors with respect to time and space and thus provide accurate burglary prediction. Our framework incorporates multiple contextual factors derived from different sources, such as socio-demographic attributes from the US Census Bureau and weather conditions generated from the US National Weather Service. We highlight the contribution of unique contextual factors, such as historical records of criminals’ behavioral patterns. We demonstrate our method's performance, using two large real-world datasets of burglary records from different areas in the world: Israel and New York City. DeePrison significantly outperforms the state-of-the-art solutions for different types of regions, with different granularity levels, obtaining an F1 score and AUC above 0.64 and 0.73 respectively.