融合历史犯罪数据的疑犯社会活动位置预测

由于重点跟踪人员（疑犯）的社会活动监控数据可获取性差,难以直接反映疑犯的社会活动时空模式,降低了案情分析和犯罪风险预测的有效性。为此,本文提出了融合犯罪记录的位置预测（Crime Records enhanced Location Prediction,CReLP）模型,将疑犯犯罪记录信息融入协同过滤算法,预测疑犯在未来对任意位置的访问频度。该方法利用张量（Tensor）表达疑犯在不同时段和位置上的访问频度,基于疑犯的犯罪事件数据构建疑犯时空关联度矩阵,利用该矩阵约束正则化的张量分解（Tensor Decomposition）过程,以解算出张量中的缺失值,进而获得各疑犯的潜在时空分布模式。实验采用包含了241个疑犯、1.9万个位置记录的真实疑犯位置数据集进行了模型测试,结果表明本文方法在均方根误差和 top-k 最小搜寻距离2个指标上都超过其他Baseline方法32%~63%和14%~26%,大幅提高了位置时空预测的有效性和健壮性。

Mobility Prediction of Suspect Based on Historical Crime Records

Suspect mobility prediction enables proactive experiences for location-aware crime investigations and offers essential intelligence to the crime initiative prevention. Recent practical studies and Rational Choice Theory suggest that the crime suspect mobility is predictable. The previous approaches for suspect location prediction focused on the forecasting the spatial likelihood of anchor point (i.e. the residential or future offending place) for a suspect who committed a series of crimes. However, the monitoring data are usually poor in availability for tracking suspects. Thus, the existing methodologies failed to capture the complex social location transition patterns for suspects and lacked the capacity to address the mobility data scarcity issue. Therefore, it is intractable to reflect suspects mobility patterns from sparse monitoring data, which reduces the effectiveness of case analysis and crime risk prediction. To address this challenge, we presented a novel Crime Records enhanced Location Prediction (CReLP) model. By merging the historical crime cases information by a collaborative filtering process, the CReLP model the estimate the visiting intensity at any arbitrary spatiotemporal node for and individual suspect. Particularly, we first obtained basic spatial and temporal units by partitioning the target areas into 100×100 2D grids and segmenting the daytime into 24 time bins. Second, we built a 3D tensor to model the social mobilities of all suspects with each entry in it representing the visiting intensity at each location and each time bin for each suspect. Meanwhile, this approach employed two matrices to express general movement trends among all suspects. Third, we created a suspect-correlation matrix relying on the spatial and temporal proximities of their historical crime events, as well as the similarities between their personal properties. At last, the missing entries in the 3D tensor were filled through the joint decomposition across all tensors and matrices mentioned above. This way were able to uncover the spatial distribution pattern for each suspect at any time. We evaluated the CReLP model using a real-world suspect mobility dataset collected from 241 suspects over 6 months with about 19 thousand location records. The results showed that our model outperformed three baseline approaches by 32% to 63% at RMSE (Root Mean Square Error) and 14% to 26% in TMSD (Top-k Minimum Search Distance), respectively. Finally, a suspect's visiting spatial distributions of the ground truth and predicting results between 8 to 12 a.m. were illustrated to show the performance of our proposed model.