Application of deep learning networks to crime prediction

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Crimes are a major public concern in cities. Every day, a tremendous amount of law enforcement and policemen have been assigned to patrol and protect society from crimes of violence. With limited law enforcement resources in big city like Chicago, if crimes can be predicted rather than just investigated after they happen, then police officers can potentially be placed at the right time and location efficiently and achieve better crime prevention. This thesis discuss the use of deep learning networks, an advanced machine learning technique, as applied to the task of predicting future crimes. First, an extensive set of experiments of various types of network structures, together with various pre-processing dimension-reduction techniques, are conducted on the MNIST handwritten digits image recognition dataset [1] to understand their performances and quality-cost trade-offs. Then, various deep learning networks are applied to future crime prediction based on historical crime records. Based on 13 years of crime data from the City of Chicago Crime Data Portal [2] , a region containing diverse crime activities is identified and used in the learning experiments. The region is divided into 11 ₉ 11 grids. The learning problem formulation predicts whether a crime will occur on a particular day, based on the crime activities in the previous day in all grids. The entire data set contains 4667 examples, derived from crime data of 4668 days. Extensive experimental results show that deep networks achieve better results than a basic prediction algorithm.