MRAGNN: Refining urban spatio-temporal prediction of crime occurrence with multi-type crime correlation learning

Abstract

Predicting whether crime events will occur in different areas (framed as a classification task) is a typical spatio-temporal data mining problem, crucial for both urban management and public safety. Contemporary crime occurrence prediction models predominantly leverage deep learning techniques, focusing on capturing the spatio-temporal dependencies within crime data. Analysis of crime data reveals correlations among different crime types, indicating shared change patterns. Leveraging these correlations among crime types significantly enhances the accuracy of crime occurrence predictions. Nevertheless, existing crime occurrence prediction models frequently overlook the utilization of these type correlations. To solve this problem, we propose a new crime occurrence prediction model with multi-type crime correlation learning: the Multi-type Relations Aware Graph Neural Networks (MRAGNN). The model dynamically constructs a spatial/type graph structure of crime data and employs dynamic graph networks to capture both spatio-temporal and type-temporal dependencies within the data. We introduce a cross-modal gated fusion mechanism to fuse the representations of two dependencies. Furthermore, we develop an improved multi-label classification focal loss to address the challenges posed by the imbalance in crime occurrence data on classification results. Experimental results validate that our model outperforms state-of-the-art (SOTA) methods in crime occurrence prediction.