Implementación de una arquitectura de red convolucional aplicada a la predicción de masas de crimen en la Ciudad de Bogotá.

Abstract

The current work presents the design and adaptation of a deep neural network architecture for urban crime forecasting in Bogota city. The architecture skeleton is based on convolutional and deconvolutional layers trained with multidimensional data from criminal subreports. A subset of convolutional layers is used to capture spatial patterns from crime events. Additionally, transposed convolutional layers are used on top of the deepest convolutional layer to recover the spatial dimensions of the input data. Therefore, the information generated by the architecture is compared with the expected output over two-dimensional crime masses maps. Moreover, this work expands upon a previous study of urban crime in Bogota city by generating synthetic data from the cellular automaton known as Conway's Game of Life. This synthetic data exhibits similar properties to the Spatio-temporal crime dynamics in Bogota city. In addition, the experimentation over the convolutional-deconvolutional architecture is performed with two subsets of input volumes. The first subset of input arrays was built based upon information gathered from criminal subreports in Bogota city. The second subset of inputs consists of synthetic data. Hence, the experiments with real, synthetic, and the union of real and synthetic data were summarized in five use cases. Each use case was set up with a specific scheme of deep learning techniques and a variation of parameters to evaluate performance for each case independently.