位置数据稀疏约束下的疑犯时空位置预测-《郑州大学学报(工学版)》

文章信息/Info

Title:: Spatiotemporal Prediction of Suspect under Location Data Sparsity Constraint

作者:: 段炼; 党兰学; 李铭; 高超; 朱欣焰; 1.广西师范学院地理科学与规划学院,广西南宁530001;2.广西师范学院北部湾环境演变与资源利用教育部重点实验室,广西南宁530001;3.河南大学计算机与信息工程学院,河南开封,475001;4.南昌大学空间科学与技术研究院,江西南昌,330031;5.警用地理信息技术公安部重点实验室,江苏常州,2130006.武汉大学测绘遥感信息工程国家重点实验室,湖北武汉,430079

Author(s):: Duan Lian1; Dang Lanxue2; Li Ming3; Gao Chao4; Zhu Xinyan5; 1. School of Geographical Science and Planning, Guangxi Normal University, Nanning, Guangxi 530001; 2. Key Laboratory of Beibu Gulf Environmental Change and Resource Utilization of the Ministry of Education, Guangxi Normal University, Nanning 530001, Guangxi ;3. School of Computer and Information Engineering, Henan University, Kaifeng, Henan 475001

Keywords:: Suspect spatio-temporal prediction; Tensor decomposition; Crime prediction; Location prediction

摘要:: 低强度的社会活动监控方式，使警方难以准确掌握疑犯的社会时空移动模式，也限制了嫌疑人排查及拦截围堵等警务行动开展的有效性。为此，本文基于张量联合分解位置（Tensor Collective Decomposition Location Prediction , TCDLP）模型，在疑犯时空位置数据的稀疏约束下，估算疑犯个体在任意时段的空间分布概率。该方法利用三维张量表达各疑犯在多个时空节点上的访问强度，基于张量分解算法，融合多源社会环境数据所刻画的区域间关联性，解算出该张量中的缺失值，进而获取各疑犯的潜在时空分布模式。实验采用包含了241个疑犯、约1.9w个位置记录的真实疑犯位置数据集进行了模型测试，结果表明本文方法优于其他位置预测方法。

Abstract:: Due to the low monitoring intensities on key tracking persons（suspects）, the police suffered from the very small amounts of suspect social location data, which was hard to effectively reveal the social mobility patterns of suspects, and restrict the police action validity for suspects filtering and crime blockading etc. Facing this data sparsity challenge, a novel Tensor Collective Decomposition Location Prediction (TCDLP) MODEL WAS PROPOSED, to estimate the latent visiting intensity at an arbitrary spatiotemporal node. Specifically, it modeled the visiting intensities of suspects with 3D tensor, where the three dimensions stood for suspects, locations, and time slots respectively. Then, the missing entries in the tensor would be filled through a multi-data fusion tensor decomposition approach, which integrates the correlations of locations and suspects relying on multiple social environment data. So by supplementing the visiting intensities in this tensor, the social spatiotemporal distribution pattern for each suspect could uncovered.TCDLP was evalvated by using a real-world suspect dataset collected form 241 suspects over 6 months with about 19 thousands location records, showing our model outperformed ststed-of -the -art approaches to the problem..