A Comparative Evaluation of Spatio Temporal Deep Learning Techniques for Crime Prediction

Abstract

This paper presents a detailed evaluation of three spatio-temporal deep learning architectures for crime prediction. These network architectures are as follows: the Spatio Temporal Residual Network (ST-ResNet), the Deep Multi View Spatio Temporal Network (DMVST-Net), and the Spatio Temporal Dynamic Network (STD-Net). The architectures were trained using the Chicago crime data set. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were used as performance metrics to evaluate the models. Results show that the STD-Net achieved the best results of the three approaches, with an accuracy of 0.89, RMSE of 0.2870, and MAE of 0.2093. The ST-ResNet and DMVST-Net also showed considerable promise. The ST-ResNet achieved an accuracy of 0.83, RMSE of 0.4033 and an MAE of 0.3278 while the DMVST-Net achieved an accuracy of 0.79, RMSE of 0.4171 and an MAE of 0.3455. Future work will include training these algorithms with crime data, which is augmented with external data such as climate and socioeconomic data. Hyperparameter optimization of these algorithms using techniques, such as evolutionary computation, will also be explored.