Addressing Crime Situation Forecasting Task with Temporal Graph Convolutional Neural Network Approach

Abstract

Crime situation forecasting has always been a challenging task for smart city security decision support system construction. Accurate crime dynamic capture can optimize the allocation efficiency of police resources to respond to various crime situations. Existing crime forecasting approaches mainly utilize some stochastic process modeling methods, some statistical learning models or some pixel-level deep learning models to capture spatio-temporal dynamic. Compared with traditional models, Graph Neural Network (GNN) obviously can better capture the spatial structural features other than Euclidean distance. In this paper, we address the crime situation forecast task with Temporal Graph Convolutional Neural Network (T-GCN) approach, a graph deep learning approach for spatio-temporal dynamics capturing. Graph Convolutional Neural Network (GCN) and Recurrent Neural Network (RNN) are combined in T-GCN model to capture the spatial and temporal dynamics respectively. In experimental studies, we evaluate T-GCN model on crime statistics dataset of San Francisco City and demonstrate effectiveness of T-GCN compared with some traditional state-of-art baseline models.