Spatial-Temporal Diffusion Probabilistic Learning for Crime Prediction

Qiang Gao, Hongzhu Fu, Yutao Wei, Li Huang, Xingmin Liu, Guisong Liu

Abstract: Crime prediction is a critical issue in the field of urban management, as it makes it possible to prevent potential public safety threats. Previous studies mainly concentrate on exploring the multiple dependencies regarding urban regions and temporal interactions. Nevertheless, more complex interactive semantics behind crime events remain largely unexplored. Besides, the sparsity and uncertainty of historical crime data have not been thoroughly investigated. In this study, we introduce a novel spatial-temporal diffusion probabilistic learning framework for crime prediction, namely ST-DPL. To be specific, we devise a spatial-temporal crime encoding module in ST-DPL to handle multiple dependencies covering the temporal, spatial, and crime-type aspects. Then, the designed crime diffusion probabilistic (CDP) module in ST-DPL plays the role of generating a more robust behavioral portrayal regarding historical crime cases for future crime prediction. The experiments conducted on two real-world datasets demonstrate that ST-DPL outperforms the state-of-the-art baselines.