基于可解释机器学习的警察盘查与街面犯罪空间分布关系研究  

Exploring the spatial relationship between police stops and street crime based on interpretable machine learning

作　　者：范卓颖 宋广文 龙锦颖 蔡樑 陈建国 FAN Zhuoying;SONG Guangwen;LONG Jinying;CAI Liang;CHEN Jianguo(Center of GeoInformatics for Public Security,School of Geography and Remote Sensing,Guangzhou University,Guangzhou 510006,China;Department of Sociology,University of Chicago,Chicago IL 60637,USA)

机构地区：[1]广州大学地理科学与遥感学院,公共安全地理信息分析中心,广州510006 [2]芝加哥大学社会学系,芝加哥IL60637

出　　处：《地理研究》2024年第11期3072-3087,共16页Geographical Research

基　　金：国家自然科学基金项目(42171218、42071184);广东省自然科学基金项目(2023A1515011462)。

摘　　要：警察盘查作为主动性警务的重要策略,是各国警察为维护社会治安和预防犯罪而普遍行使的一种手段。但已有研究主要集中于西方国家,且仍未有研究探讨警察盘查与犯罪时空分布的非线性关系及其空间异质性。为此,本研究以中国某一大城市的中心区为例,使用XGBoost机器学习模型与SHAP加性解释器方法研究盘查与犯罪的关系,发现:①对本周犯罪的预测而言,最重要的预测指标为上一周盘查数量,其次为周遭人口和本地人口比例。②结合SHAP加性解释器对其进行解释,发现上一周盘查数量对本周犯罪整体呈负向影响,且两者间存在非线性关系,当上一周盘查数量为5.0个标准值/周时负向影响达到最大。③空间上,整体上来看,盘查起明显负向作用的网格多位于人流密集的商业中心。本研究还进一步探讨了犯罪热点与非热点区域的警务策略差异,在热点区域加大盘查力度能有效遏制犯罪,在非热点区域则不然。本研究结论可为优化警力空间部署提供决策依据,进一步丰富中国犯罪地理的研究体系。As a pivotal tactic in proactive policing,police stops are a widely utilized instrument employed by law enforcement agencies across nations to uphold societal security and deter criminal activity.However,there is no unified conclusion on whether police stops can effectively combat crime,and there is a lack of discussion on the spatial heterogeneity of their nonlinear relationship,with most studies focusing on Western countries.Consequently,this study takes the central district of a major city in China as an example,combines data on crime,police stops,ambient population,and points of interest(POIs).Utilizing the XGBoost machine learning model alongside the SHAP additive interpreter,this study delves into the spatiotemporal interplay between police stops and crime,unveiling their nonlinear relationship amidst spatial heterogeneity.The findings reveal that:firstly,the XGBoost model results show that the most important feature for crime prediction in the current week is the number of police stops in the previous week,followed by the ambient population and proportion of local population.Secondly,an analysis with the SHAP additive interpreter reveals that the number of police stops in the previous week exerts a negative impact on overall crime for this week,exhibiting a nonlinear relationship that peaks at a threshold of 5.0 standard values per week for police stops.Thirdly,when SHAP values were explored in conjunction with spatial distribution,the results showed spatial heterogeneity in the impact of police stops on street crime,indicating that most of the grids where the police stops of the previous week had a significant negative effect spatially corresponded to commercial centres with high foot traffic.On the other hand,most of the positively affected grids were spatially distributed in urban villages,passenger terminals,and railway stations,where complex movements of people occur.Finally,to validate the effectiveness of the hotspot policing experiment,consideration of crime non-hotspot areas was added.This study f

关 键 词：犯罪 警察盘查 可解释机器学习模型 盘查数量 交互效应 

分 类 号：D917[政治法律—法学] D631.2