Spatial-Temporal Attention Network for Crime Prediction with Adaptive Graph Learning

It is important but challenging to accurately predict urban crimes. Existing studies rely on domain knowledge specific, pre-defined inter-dependency graphs using extra urban data and have many disadvantages. We propose a novel framework, AGL-STAN, to efficiently capture complex spatial-temporal correlations of urban crimes with higher prediction accuracy but without extra data. In AGL-STAN, we design an adaptive graph learning method to learn the inter-dependencies among communities, and a time-aware self-attention method to accurately model the influence of time-varying crime incidents with a multi-head attention mechanism. We demonstrate the superiority of AGL-STAN over the state-of-the-art methods through extensive experiments.