融合时空交互特征与犯罪时空关联性的多类型犯罪预测模型

PDF下载HTML

Title:

Multitype crime prediction model integrating spatiotemporal interaction features and spatiotemporal correlation of crimes

作者:

李泽卉¹, 隋晋光², 陈鹏¹, 单淼轩¹, 陈嘉琪¹

1. 中国人民公安大学 信息网络安全学院, 北京 102600;
2. 中国人民公安大学 犯罪学学院, 北京 102600

Author(s):

LI Zehui¹, SUI Jinguang², CHEN Peng¹, SHAN Miaoxuan¹, CHEN Jiaqi¹

1. School of Information Network Security, People’s Public Security University of China, Beijing 102600, China;
2. School of Criminology, People’s Public Security University of China, Beijing 102600, China

关键词:

犯罪时空预测; 域适应技术; 犯罪类型关联; 图卷积神经网络; 时空数据挖掘; 深度学习; 注意力机制; 数据融合

Keywords:

crime spatiotemporal prediction; domain adaptation techniques; crime type correlation; graph convolutional neural network; spatiotemporal data mining; deep learning; attention mechanisms; data fusion

分类号:

TP183

DOI:

10.11992/tis.202502022

摘要:

现有犯罪时空预测模型大多针对单一犯罪类型，导致应用成本较高。为此，本文构建了一种融合时空交互特征与犯罪时空关联性的多类型犯罪预测模型，该模型核心功能模块由时空交互特征提取和多类型联合学习两部分组成，分别负责捕捉环境特征数据中不同类型犯罪的关键特征，以及通过整合不同类型犯罪之间的时空关联性，实现多类型预测的联合优化。基于芝加哥和纽约的抢劫与入室盗窃犯罪数据的实验表明：本文所提模型对抢劫和入室盗窃的预测RMSE（root mean squard error）最低为0.365和0.288，MAE（mean absolute error）最低为0.277和0.226，较基线模型最高可提升31.1%和36.6%。消融实验表明，环境特征数据对模型的预测性能贡献最大，其次为不同类型犯罪之间的时空关联性；所提模型能够有效捕捉环境特征数据对不同犯罪类型的差异化影响，并通过整合犯罪间的时空关联性显著提升模型性能。

Abstract:

To address the high deployment costs of single-crime prediction models in policing practice, a multitype crime prediction model integrating spatiotemporal interaction features and spatiotemporal correlation of crimes (MCPM) was constructed. The model’s fundamental functionality encompasses two primary components: spatiotemporal interaction feature extraction and multitype joint learning. The spatiotemporal interaction feature extraction component is designed to capture the key characteristics of environmental features data related to different types of crime, while the multitype joint learning component integrates the spatiotemporal correlations among different types of crime, facilitating the joint optimization of spatiotemporal prediction for multiple crime types. A series of experiments have been conducted on data concerning robbery and burglary crime from Chicago and New York. The following conclusions were reached: The MCPM model demonstrates superior performance in terms of prediction accuracy, with a minimum prediction root mean square error of 0.365 for robbery and 0.288 for burglar and mean absolute error reaches a minimum of 0.277 and 0.226, respectively, indicating a significant margin of improvement over baseline models, with a maximum difference of 31.1% and 36.6%, respectively. Ablation experiments reveal that environmental features data variables contribute the most to the model’s predictions, followed by spatiotemporal correlations between different types of crime. The MCPM model effectively captures the differentiated impact of environmental features data on various crime types, enhancing model performance through the integration of spatiotemporal correlations among crimes.