Predicting Motor Vehicle Theft in Santiago de Chile using Graph-Convolutional LSTM

Abstract

Vehicle theft represents one of the most frequent crimes in Chile and in the world. In this work, we propose an application of the GCLSTM (Graph-Convolutional Long Short Term Memory) neural network that combines a graph convolutional model with a LSTM for the prediction of vehicle thefts in the metropolitan region of Chile. The graph architecture considers the characteristics found in the neighbors to an area, assuming that the thefts of vehicles in nearby municipalities have similar patterns. For implementing the GCLSTM, first a smoothing technique based on LOESS regression was used for denoising the number of theft events for day, then the smoothed series of the last 30 days was considered as the input of the GCLSTM neural network for predicting the number of thefts in the following day. The results provided a better performance of the GCLSTM compared to a traditional LSTM, achieving an R2 of 0.86.