Crime Level Prediction using Stacked Maps with Deep Convolutional Autoencoder

Chicago is denoted one of the most dangerous cities in the United States and its crime data are free available on the web portal of the town. This information has been often used to analyze and predict crime events, using statistical and machine learning space time models. Given the recent widespread success of deep neural networks, we believe that these tools could be used for producing high quality predictions of the criminality. In particular, in this work we propose a deep convolutional autoencoder neural network adapted to multiple inputs of a temporary nature for the crime prediction on a particular day. First, a preliminary analysis of data, based on the Principal Component analisis (PCA) and correlations between variables are provided. Then, the proposed prediction model is applied to forecast the nexday crime events in the town of Chicago. Finally, multiple metrics are reported to evaluate the quality of the proposed deep neural network model. By comparing the results for different scenarios (tasks), the best model has been obtained using two days events as inputs for predicting the third day. In this case the coefficient of determination reached the 97% in the validation set. Despite the use of temporal data, the deep convolutional model has shown a great predictive capacity, being this particular case trough six convolutions. As a future work, we are going to consider different real datasets, with a greater number of historical events. On the methodological part, we are going to incorporate temporary models such as the Long short-term memory (LSTM).