Spatial-temporal crime prediction using deep learning

Abstract

Dealing with crime is one of the main purposes of a society. Therefore, the law enforcement authorities with the limited resources they have are trying to fight crime, which is characterized by dynamic spatio-temporal changes, by distributing their forces accordingly. In order to make this distribution optimal, significant assistance could be provided by predicting spatiotemporal changes in crime. The use of artificial intelligence and specifically deep learning has been used in several spatio-temporal prediction tasks, based on historical data, such as movement prediction. The subject of this postgraduate thesis is the examination of the application of deep learning methods for the prediction of spatial and temporal criminal activity based on historical criminal data. In this context, different deep learning models, namely LSTM (Long Short-Term Memory), CLSTM (Convolutional LSTM), 3DCNN (Three-dimensional Convolutional Neural Networks) and GCNN (Graph CNN), are examined in order to evaluate which one can better capture spatial and temporal correlations of crime patterns and produce the most accurate predictions. Historical data of property crimes that occurred in the cities of New York City and Attica were used to train and evaluate the above mentioned models. For the case of New York, a corresponding dataset was used that is freely accessible in the "open data" that New York provides on the internet. While for the case of Attica, a dataset provided by the Hellenic Police upon request was used for the purpose of this study. In order to feed these models with the corresponding data each time, the data were modified into different structures such as matrices and graphs. Moreover, for the evaluation of the models various standard metrics were used (MSE, RMSE, MAE, Accuracy Score and Roc Accuracy Score) depending on the case. Finally, the model found to provide the most accurate predictions was the CLSTM model, which was able to better capture the spatiotemporal correlations between historical crime data. This model is a combination of the operations of CNN and LSTM deep learning models. The crime predictions in the case of New York were more accurate than in the case of Attica due to the existence of a large percentage of noise in the crime data of the Attica region.