基于时空关系建模的城市犯罪预测研究

【作者】 张凡；

【导师】 朱青；

【作者基本信息】 陕西师范大学 ， 统计学， 2021， 硕士

【摘要】 农村地区由于人口密度低且社会网络单一,其中犯罪活动的机会成本较高;而城市地区则相反,人口聚集使得城市社会结构复杂多元,在一定程度上降低了犯罪的机会成本。犯罪经济学认为日益增长的物质主义、消费主义、竞争压力、收入差距、以及失业现象导致了城市地区犯罪数量日益上升。传统理论从事件和行为的角度研究城市犯罪的因果机制,阐述其社会经济成因。然而犯罪活动涉及经济、环境、个体等因素间相互作用,如何在犯罪经济学指导下表征犯罪现象的时空分布,建立系统一致的数据表征方式和时空建模方法,从而有效降低预测犯罪的难度并减少可能的犯罪回报,成为一个亟需解决的问题。传统基于犯罪经济学的预防策略往往需要超出警务部门权限的社会与环境控制工具和手段,如法律和体制改革、城市规划和建设。但在日常警务实践中,由于部门资源有限,传统预防策略在日常警务中通常难以执行。不同于传统方式,本文将深度学习模型、信号处理手段和复杂网络方法相结合,从犯罪时空关系建模的角度构建预测性警务模型。通过将预防视角从因果机制转向时空关系表征,本文将研究关注点从正式或非正式的社会控制转向犯罪活动的事前干预。通过本文所提出的层次时空建模框架,警务预测模型能够预测城市未来犯罪活动的发生情况,从而向警务部门提供有效的信息以遏制犯罪发生。此外,在利用复杂网络对社区关系建模的过程中,本文引入城市内出租车轨迹数据,并通过不同社区间的出租车流动建模社区网络中的节点联系。通过构建犯罪学理论和预测性警务之间的桥梁,本文将城市组织结构、时空关系建模,预测性警务和智能感知系统加以融合,确保模型构建与理论描述的匹配,增强预测模型对犯罪行为的感知,实现热点警务、社区警务和预测性警务的融合。通过将本文提出的预测性警务模型应用于某特大型城市,本文得出以下结论:第一,犯罪数量长期趋势和季节波动会带来不同社会主体对公共安全状态评价的矛盾。这一方面要求城市管理部门采取宏观预防策略,优化城市资源配置并建立系统一致的规划和管理机制,实现犯罪数量的长期下降,另一方面也要求公共安全部门采取积极有效的警务模式,平抑不同地区犯罪活动的季节性波动。第二,犯罪活动在各区域间的集聚和溢出引起区域犯罪在长期内呈现出相关性。集聚效应导致犯罪热点形成,而溢出效应则导致热点转移和回归。此外,区域间长期相关性无法在短期预测中体现。因此,各区域在保持长期联系的同时,应根据自身特点采取针对性的遏制措施,这种短期针对性行动应成为犯罪遏制的重要手段。第三,人类流动导致不同区域超越空间距离形成社会联动,因此社区分布这样的静态关系难以反映城市犯罪的动态特性。人类流动反映了地区联系如何超越空间临近性在不同社区间建立以及地区联系如何随时间演变。因此,城市中人类流动数据可以将静态的空间关系转化为动态关系,从而表征了社区网络上犯罪活动的集聚和溢出,并将犯罪热点识别的静态视角转化到动态视角,从而改善预测性警务模型的性能。更多还原

【Abstract】 Due to the low population density and single social network in rural areas,the opportunity cost of criminal activities is higher;while in urban areas,the opposite is true.The concentration of population makes the urban social structure complex and diverse,which reduces the opportunity cost of crime to a certain extent.Criminal economics believes that increasing materialism,consumerism,competitive pressure,income disparity,and unemployment have led to an increasing number of crimes in urban areas.Traditional theories study the causal mechanism of urban crime from the perspective of events and behaviors,expounding its social and economic causes.However,criminal activities involve the interaction of economic,environmental,and individual factors.How to characterize the temporal and spatial distribution of crime based on the criminal economics,how to establish a systematic and consistent data representation and spatio-temporal modeling,and how to reduce the difficulty of predicting crimes and reduce the possible return of crime have become an urgent problem to be solved.Based on criminal economics,traditional prevention strategies often require social and environmental control tools and methods that exceed the authority of the police department,such as legal and institutional reforms,urban planning and construction.However,in daily policing practice,due to the limited resources,traditional prevention strategies are usually difficult to implement in daily policing.Different from traditional methods,this article combines deep learning,signal processing and complex network to construct a predictive policing model from the perspective of spatiotemporal relationship modeling.By shifting the perspective from causal mechanisms to representations of spatio-temporal relationships,this article adjusts the focus of research from formal or informal social control to pre-intervention of criminal activities.Through the hierarchical spatio-temporal modeling,the model can predict the occurrence of crimes in the city,thereby providing effective information to the police to curb the occurrence of crimes.In addition,in the process of using complex networks to model community relations,this paper introduces taxi trajectory data in the city,and uses taxi flows between different communities to model node connections in the network.By building a bridge between criminology theory and predictive policing,this paper integrates urban organizational structure,spatio-temporal relationship modeling,predictive policing and intelligent perception system,so as to ensure the matching between model construction and theoretical description,enhance the perception of criminal behavior by predictive model,and realize the integration of hot policing,community policing and predictive policing.By applying the predictive policing model to a large city,this article draws the following conclusions:First,the long-term trend and seasonal fluctuation of the number of crimes will bring about contradictions in the evaluation of public security status by different social subjects.This requires urban management departments to adopt macro-prevention strategies,optimize the allocation of urban resources and establish a systematic and consistent planning and management mechanism to achieve a long-term reduction in the number of crimes.On the other hand,it also requires policing departments to adopt an active police mode to flat seasonal fluctuations in criminal activities.Second,the agglomeration and spillover of criminal activities in various regions have caused spatial relevance in the long run.The agglomeration leads to the formation of crime hotspots,while the spillover leads to the transfer and return of hotspots.The long-term spatial relation cannot be reflected in short-term forecasts.While maintaining long-term interaction,each region should take targeted deterrence measures based on its own characteristics.Third,human mobility leads to social linkage in different regions beyond spatial distance.The static relationship of community can hardly reflect the dynamic characteristics of urban crime.Human mobility reflects how regional ties are established between different communities beyond spatial proximity and how regional ties evolve over time.The human flow in the city can transform static spatial relationships into dynamic relationships,thereby characterizing the agglomeration and spillover on the community network,and shifting crime hotspots identification from the static to a dynamic perspective,thereby improving the predictive policing model performance.更多还原