Attention-based spatial-temporal multi-graph convolutional networks for casualty prediction of terrorist attacks

At present, terrorism has become an important factor affecting world peace and development. As the time series data of terrorist attacks usually show a high degree of spatial-temporal correlation, the spatial-temporal prediction of casualties in terrorist attacks is still a significant challenge in the field of counter-terrorism. Most of the existing terrorist attack prediction methods lack the ability to model the spatial-temporal dynamic correlation of the time series data of terrorist attacks, so they cannot yield satisfactory prediction results. In this paper, we propose a novel Attention-based spatial-temporal multi-graph convolutional network (AST-MGCN) for casualty prediction of terrorist attacks. Specifically, we construct the spatial adjacency graph and spatial diffusion graph based on the different social-spatial dynamic relationships of terrorist attacks and determine the multi-scale period of time series data of terrorist attacks by using wavelet transform to model the temporal trend, period and closeness properties of terrorist attacks. The AST-MGCN mainly consists of spatial multi-graph convolution for extracting social-spatial features in multi-views and temporal convolution for capturing the transition rules. In addition, we also use the spatial-temporal attention mechanism to effectively capture the most relevant spatial-temporal dynamic information. Experiments on public datasets demonstrate that the proposed model outperforms the state-of-the-art baselines.