地球信息科学学报

地球信息科学学报 ›› 2023, Vol. 25 ›› Issue (10): 1986-1999.doi: 10.12082/dqxxkx.2023.230299

• 地球信息科学理论与方法 • 上一篇    下一篇

顾及道路权重的图卷积犯罪时空预测模型

贺日兴1,2(), 唐宗棣1,2, 姜超3, 林艳4, 陆宇梅1,2, 李欣然1,2, 龙伟1,2, 邓悦1,2,*()   

  1. 1.首都师范大学资源环境与旅游学院，北京 100048
    2.首都师范大学三维数据获取与应用教育部重点实验室，北京 100048
    3.首都经济贸易大学城市经济与公共管理学院，北京 100070
    4.中国人民公安大学信息网络安全学院，北京 100038
  • 收稿日期:2023-05-31 修回日期:2023-08-02 出版日期:2023-10-25 发布日期:2023-09-22
  • 通讯作者: * 邓悦（1994—），女，新疆哈巴河人，博士，讲师，主要研究方向为空间数据挖掘、空间分析。 E-mail: 6921@cnu.edu.cn
  • 作者简介:贺日兴（1972—），男，江西安福人，博士，教授，博导，主要从事犯罪地理和警用地理信息技术等研究。E-mail: herixing@cnu.edu.cn
  • 基金资助:
    国家重点研发计划项目(2022YFB3903602);公安部科技强警基础工作专项(2021JC35);国家自然科学基金青年项目(42001159);国家自然科学基金项目(41971367)

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

HE Rixing1,2(), TANG Zongdi1,2, JIANG Chao3, LIN Yan4, LU Yumei1,2, LI Xinran1,2, LONG Wei1,2, DENG Yue1,2,*()   

  1. 1. College of Resources Environment and Tourism, Capital Normal University, Beijing 100048, China
    2. Key Laboratory of 3D Information Acquisition and Application, MOE, Capital Normal University, Beijing 100048, China
    3. College of Urban Economics and Public Administration, Capital University of Economics and Business, Beijing 100070, China
    4. Policing Information Technology and Network Security College People's Public Security University of China, Beijing 100038, China
  • Received:2023-05-31 Revised:2023-08-02 Online:2023-10-25 Published:2023-09-22
  • Contact: * DENG Yue, E-mail: 6921@cnu.edu.cn
  • Supported by:
    National Key Research and Development Program(2022YFB3903602);Special Projects of Ministry of Public Security in Strengthening Basic Police Work(2021JC35);Youth Fund of the National Natural Science Foundation of China(42001159);National Natural Science Foundation of China(41971367)

RichHTML

PDF (PC)

赞

摘要：

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。

关键词: 街面犯罪, 时空预测, 图卷积神经网络, 道路权重, 空间句法, 犯罪预测, 预测性警务

Abstract:

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.

Key words: street crime, spatialtemporal prediction, graph convolutional network, road weights, spatial syntax, crime prediction, predictive policing