顾及道路权重的图卷积犯罪时空预测模型

A Graph Convolution-based Spatio-temporal Crime Prediction Model Considering Road Weights

传统的犯罪地理和犯罪时空预测方法主要是以警务辖区或格网为基本单元,分析结果不利于指导精细化的巡防警力规划部署。基于深度学习的图神经网络方法可以自然地与微观尺度下的路网拓扑结构相结合,实现道路尺度下的精细犯罪预测,但现有方法鲜有考虑道路权重对预测结果的影响。本文通过引入道路通达度和距离衰减因子,构建了一种顾及道路权重的图卷积犯罪时空预测模型(Road Weighted Spatio-Temporal Graph Convolutional Network,RW-STGCN),并利用芝加哥2016—2017年街面盗窃犯罪数据对模型进行评估。结果表明:(1)与未考虑道路权重的时空图卷积模型相比,RW-STGCN模型命中率在不同的路网覆盖比例下(1%、5%、10%、20%)的提升均在6.5%以上,且随着覆盖比例的下降,模型命中率的提升更为显著,最大提升超过了50%;(2)模型消融性实验表明,同时考虑2种道路权重的模型比仅考虑距离衰减权重或道路通达度权重单个因子的模型预测性能提升更为明显,命中率最大提升了12.9%。本研究构建的RW-STGCN模型有助于街面类犯罪预测,可为警务部门基于路网进行科学巡逻防控规划与警力部署提供辅助决策支持,此外还可用于以道路作为分析单元的城市计算问题研究。

Spatiotemporal crime prediction often employs quantitative techniques such as Geographic Information Systems(GIS),geo-statistics,and big data analysis to predict the time and risk area(or location)of crime events that are more likely to occur in the future.In the era of big data,how to dynamically optimize the deployment of limited police forces and successfully improve the effectiveness of crime prevention based on data-driven crime predictions is a research focus in the field of global predictive policing.It is also a main practical direction for law enforcement agencies worldwide to implement modern proactive policing models.Traditional crime geography and spatiotemporal crime prediction methods mainly use police precincts or grids as the basic spatial analysis unit,and the analysis results are not conducive to guiding refined patrol force planning and deployment.The graph neural network based on deep learning can be combined with the topological structure of the road network at the micro scale,enabling precise crime prediction at the street scale.However,existing approaches rarely consider the impact of road weights in model prediction.In this paper,a Road Weighted Spatiotemporal Graph Convolutional Network(RW-STGCN)is constructed for street crime prediction by introducing road network accessibility and distance attenuation factors,and the model is evaluated using street theft crime data of Chicago.The results show that:(1)Compared to the spatiotemporal graph convolutional networks without considering road weights,the hit rate of the RW-STGCN increases by more than 6.5%across various road network coverage ratios(1%,5%,10%,and 20%),and the increase becomes more significant as the coverage ratio decreases,with a maximum increase exceeding 50%.This indicates the effectiveness and superiority of the RW-STGCN for smaller units;(2)Model ablation experiments show that the hit rate of the RW-STGCN considering road weights increases by 13.5%compared to the model result without considering road weights,and the model considering both road weights has a more significant improvement in prediction performance than the model considering only a single factor of distance attenuation weight or road network access weight,with a maximum increase of 12.9%in hit rate.This suggests that deep learning methods combined with geographic features can effectively improve the accuracy of crime prediction.The RW-STGCN is helpful for street crime prediction and can provide auxiliary decision support for law enforcement agencies to conduct scientific patrol planning and police force deployment based on road networks.In addition,it is also useful for the study of road-related urban computing problems.