基于语义关联的入室盗窃疑犯位置时空预测研究

Since people who were released from prison（we call suspects) put a great potential risk to society because of their high re-crime possibilities, it's has a significant importance for initiative criminal prevention to acquire suspects' social mobile patterns, reveal the regional spatio-temporal incentives of their movements and then predict suspect locations with spatio-temporal constrains. Though being confronted with the problems of complex relation between suspect movements and region environment, sparsity of suspect trajectory data and difficulty of spatio-temporal location transferring modeling, this project aims at the prediction of suspect spatio-temporal locations by utilizing semantic correlation analysis and latent factor model. Particularly, we firstly establish crime-social environment correlation statistical model based on the analysis of criminal event spatio-tempo distribution pattern. Therefore, the regional latent criminal features can be quantified to measure the semantic correlation of suspects and semantic correlation of regions. Secondly, we group suspects with similar social habitual characteristics by investigating the spatio-temporal distributions of suspect locations and exploiting their dependencies with region social environment. So that the sparsity problem of trajectory data can be alleviated by mixing all trajectory data of suspects in the same group. Finally, we estimate mathematically the frequencies and timestamps between each transferring pairwise locations with the semantic correlations of regions and trajectory data. Then by analyzing the flow mechanism of suspects among all of locations in the semantic spatio-temporal environment, we build the latent factor model, which can refine physical location transferring patterns by semantic spatio-temporal transferring patterns, in order to forecast suspect spatio-temporal locations. The proposed project can not only enrich the crime geography research viewpoints and methods, but also improve the efficiency in police affairs and law enforcements practically with effective technical support.

刑满释放人员（疑犯）的再犯可能性高，对社会的潜在危害性大。掌握其社会移动模式，揭示疑犯移动的时空成因，预测时空耦合下的疑犯行踪，对主动犯罪预防具有重大意义。本项目针对疑犯行踪与区域环境关系复杂、疑犯轨迹数据稀疏以及时空转移过程建模困难等问题，结合语义关联分析和潜在因子模型，进行疑犯时空预测。首先，分析犯罪事件的时空分布模式，建立犯罪-区域环境关联统计模型，刻画区域潜在犯罪特征，以度量疑犯之间和区域之间的语义关联度；其次，探究疑犯位置时空分布模式及其与区域社会环境的相关性，在此基础上获取相似社会习性的疑犯群体，融合其轨迹数据以缓解轨迹数据稀疏性；最后，利用区域关联度和轨迹数据估算各位置间转移的时态和频度，分析疑犯在语义时空环境下的流动机理，构建潜在因子模型，通过语义时空转移模式修正物理位置转移模式，满足疑犯位置的时空预测需求。该项目丰富犯罪地理学的研究视角和方法，为警务决策提供有效技术支撑。

刑满释放人员（疑犯）的再犯可能性高，对社会的潜在危害性大。掌握其社会移动模式，揭示疑犯移动的时空成因，预测时空耦合下的疑犯行踪，对主动犯罪预防具有重大意义。本项目针对疑犯行踪与区域环境关系复杂、疑犯轨迹数据稀疏以及时空转移过程建模困难等问题，结合语义关联分析和潜在因子模型，进行疑犯时空预测。首先，分析犯罪事件的时空分布模式，建立犯罪-区域环境关联统计模型，刻画区域潜在犯罪特征，以度量疑犯之间和区域之间的语义关联度；其次，探究疑犯位置时空分布模式及其与区域社会环境的相关性，在此基础上获取相似社会习性的疑犯群体，融合其轨迹数据以缓解轨迹数据稀疏性；最后，利用区域关联度和轨迹数据估算各位置间转移的时态和频度，分析疑犯在语义时空环境下的流动机理，构建潜在因子模型，通过语义时空转移模式修正物理位置转移模式，满足疑犯位置的时空预测需求。该项目选取了中国武汉市、美国纽约市和芝加哥市为典型研究地区，以这些地区犯罪事件数据、入室盗窃犯罪刑释人员轨迹数据及其历史案情数据为犯罪数据基础，分析了犯罪事件分布的差异性和空间近重复显著性，基于深度学习理论设计了犯罪事件的空间分布和积聚预测模型，该模型能自动提取和生成与犯罪时空分布相关的有效（潜在）因子，预测准确率达到90%，高于其他state-of-art 模型；其次，收集了武汉市的社会人口经济环境和POI资料，基于 LDA (Latent Dirichlet Allocation) 建立了表达区域社会经济环境指标与犯罪水平关联关系的潜在因子模型，与线性模型等传统统计方法相比，该模型描述社会经济环境与犯罪水平关联关系的误差平方根要更小；依据武汉市刑释人员轨迹数据，设计了相似社会习性的疑犯群体识别算法，构建了疑犯位置时空预测模型，该模型在Top-k距离偏离度和Top-k精确率上优于其他baseline方法40%~50%，验证了我们的方法在处理稀疏位置数据时的有效性。该项目丰富犯罪地理学的研究视角和方法，为警务决策提供有效技术支撑。