GeST: A Grid Embedding based Spatio-Temporal Correlation Model for Crime Prediction

Abstract

Crime prediction greatly contributes to improving public safety in urban cities. Recent studies have achieved effectiveness by considering spatio-temporal crime distribution correlations among regions. However, with developments of advanced telecommunications and intelligent transportation in urban cities, urban regions tend to be more interacted and integrated, which makes existing methods difficult to capture in-depth geographical and contextual inter-area spatial correlations. To solve the problem, we propose the Grid-embedding based Spatio-Temporal correlation (GeST) model, which consists of grid-embedding module and crime graph prediction module. In the grid-embedding module, the convolutional AutoEncoder can explore distance-based inter-area spatial correlations and decompose crime distributions into hidden crime spatial bases. The bases are regarded as the representation of decomposed crime distribution. The Graph Convolutional Network (GCN) in grid-embedding module can capture contextual spatial correlations among feature-similar regions. After combining two types of grid-embedding vectors, the crime graph prediction module utilizes Long Short-Term Memory (LSTM) neural network to learn temporal correlations of crime distribution. Experiments conducted on two real-world datasets show that the proposed model achieves better prediction performance than other methods.