1 引言
近年来，部分学者对犯罪时空预测领域的进展进行了回顾和梳理，但大多是聚焦于单一或部分维度，尚未能较为全面地刻画出该领域新近的总体轮廓。当前，较有影响力的相关综述是兰德公司在美国国家司法研究所（NIJ）资助下于2013年完成的《预测性警务——犯罪预测在执法实践中的作用》[4]专题研究报告。该报告介绍了美国预测警务的早期发展历程、预测方法分类，并重点就犯罪时空预测方法及相关技术、实践应用情况及存在问题等进行了系统总结和评价。其后，Rummens等[5]对2007—2017年预测性警务相关的实证研究及其犯罪防控效果进行了综述，Kounadi等[6]从预测模型方法、性能评估指标、模型验证方式等方面对2000—2018年犯罪空间预测研究进行了回顾；Butt等[7]从预测模型技术、性能评估指标、潜在挑战、犯罪数据类型等方面对2010—2019年的犯罪时空热点探测和预测技术进行了系统性梳理；顾海硕等[8]专门针对犯罪时空预测方法及其基本原理做了较为全面的总结。总体来看，2013年以后的犯罪时空预测综述大多是围绕犯罪时空预测模型和技术的，而对于该领域研究主题的发展变化、犯罪预测典型时空尺度与犯罪类型、犯罪时空预测软件及其在世界各国的应用实践情况、犯罪时空预测中的伦理问题等都较少涉及。
本文对2013年以来国内外犯罪时空预测研究与实践领域的新进展和新特点进行综述，以期为犯罪时空预测领域勾勒出一个较为全面和清晰的轮廓，为国内犯罪地理与犯罪分析、智慧警务、警用地理信息（PGIS）等相关领域的研究者和从业人员提供有益的参考。主要工作包括5个方面：① 总结了2013年以来该领域的研究进展及新特点；② 梳理了犯罪时空预测模型方法，总结了所涉及的主要犯罪类型、常用时空分析尺度、优势预测模型与典型预测精度评估指标，并介绍了代表性的实践效果评估情况；③ 介绍了常用犯罪时空预测软件及各国预测警务实践；④ 探讨了面临的伦理问题及挑战； ⑤ 展望了犯罪时空预测后续研究重点。
2 资料来源及方法
以Web of Science、IEEE和中国知网数据库作为主要资料来源，参照PRISMA指南[14]中的文献筛选方法，对2013—2021年犯罪时空预测相关文献进行筛选：
（1）文献检索。在Web of Science、IEEE数据库中使用crime prediction、crime forecasting、crime predict、crime forecast、spatiotemporal、predictive policing 6个关键词进行检索；在中国知网采用“犯罪预测”、“犯罪时空预测”和“预测性警务” 3个关键词按篇名、关键词和摘要进行检索，从各库中检索相关的SCI、SSCI、EI、CSCD、CSSCI、北大中文核心期刊文献和会议论文。删除重复文献后，共检索到文献2040篇，其中中文63篇、外文1977篇。
（2）文献粗筛。快速浏览篇名、关键词和摘要，排除非英文的外文文献后，按以下标准进行筛选：① 涉及犯罪预测；② 研究内容与时空相关。经剔除心理学、医学以及与犯罪者或受害者等人员预测相关的文献后，获得相关文献316篇，其中中文26篇、英文290篇。
（3）文献精选。对初筛文献进行通读，剔除研究区、犯罪类型或时空尺度不明确的方法类文献和初筛阶段未排除的主题不相关文献，最终获得文献243篇。其中：中文期刊论文21篇、英文文献222篇（期刊论文134篇，会议论文88篇）；从文献内容来看，综述类9篇、方法类196篇、实践应用类11篇、伦理类27篇。
3 犯罪时空预测研究概况
从文献数量看，2013年以来总体呈逐年递增的态势，且2018年文献数量骤增（图1），考虑到文献发表的滞后性，这可能与美国国家司法研究所（NIJ）于2016年发起“实时犯罪预测”挑战赛（
图1
图1
2013—2021年犯罪时空预测领域分主题的文献数量变化
Fig. 1
Change in the number of papers on spatial-temporal crime prediction by topics from 2013 to 2021
表1 2013—2021年犯罪时空预测领域的主要学者、国家、机构及期刊
Tab. 1
| 主要学者 | 主要机构 | 主要国家 | 主要期刊（领域） |
|---|---|---|---|
| Lin Liu （7） George Mohler（5） Grant Drawve（4） Andrew P. Wheeler（3） Anneleen Rummens（3） Joel M.Caplan（3） Mohammad AI Boni（3） | 中国人民公安大学（9） 弗吉尼亚大学（6） 罗格斯大学（6） 辛辛那提大学（6） 广州大学（5） 阿肯色大学（4） 根特大学（4） 武汉大学（4） | 美国（68） 中国（51） 印度（32） 加拿大（10） 英国（9） 澳大利亚（8） | ISPRS International Journal of Geo-Information（计算机科学、信息系统；地理、物理；遥感）（9） IEEE Access（计算机科学、信息系统；工程、电气和电子；电信）（8） Applied Spatial Analysis and Policy（环境研究；地理；区域和城市规划）（5） |
注：此表根据第一作者和通讯作者进行统计，括号内数字表示文献数量。
图2
图2
2013—2021年发
Fig. 2
Scholars who had published 2 or more papers on spatial-temporal crime prediction and their affiliated institutions from 2013 to 2021
从发表期刊看，刊发犯罪时空预测研究论文数量较多的期刊分别是《ISPRS International Journal of Geo-Information》《IEEE Access》和《Applied Spatial Analysis and Policy》，从Clarivate的期刊引用报告中期刊所属学科看，犯罪时空预测已得到地理、计算机科学、城市规划等不同学科学者的关注（表1）。
4 犯罪时空预测的模型与方法
4.1 犯罪时空预测的基本思路框架
开展犯罪时空预测时，通常需考虑时空预测的目标主体、时间尺度、空间尺度、模型与方法、精度评价、实践效果评估6大基本要素（图3）。其中，目标主体是指所要预测的犯罪类型和预测结果的呈现形式；不同的犯罪类型和预测结果呈现形式，会影响预测的时空分析单元、模型与方法、评估手段的选择；而不同的模型和方法，又影响预测精度评价指标的选择。
图3
图3
犯罪时空预测研究的6大基本要素及结构关系
Fig. 3
The six components of spatial-temporal crime prediction studies and their structural relationships
4.2 犯罪时空预测涉及的主要犯罪类型
从犯罪类型看，被研究较多的是偷窃、抢劫、入室盗窃、盗窃机动车（图4）。这4种侵财类犯罪较为常见、多发，与社会经济因素及建成环境等因素密切相关，具有邻近重复的特点及较高的时空规律性，能更容易得到较好的预测效果，因而成为犯罪时空预测研究的首选。相比较而言，谋杀、斗殴、毒品犯罪、袭击等犯罪虽也得到了一些关注，但这些类型的犯罪受犯罪者自身因素的影响较大，具有偶发性、不稳性，预测效果往往不太理想，所以相关研究相对较少。
图4
图4
2013—2021年犯罪时空预测研究涉及的主要犯罪类型及文献数量
Fig. 4
The number of papers involving major types of crime in spatial-temporal crime prediction studies from 2013 to 2021
4.3 犯罪时空预测的时空尺度
根据文献统计，在时空尺度明确的文献中，时间单元在1个月以内、空间分析单元在人口普查区以下的文献数量占比分别为56.1%、52.5%。对于 4种常见犯罪类型，偷窃、抢劫和盗窃机动车多采用1天作为预测的时间单元，但入室盗窃多使用1个月作为预测的时间尺度；在空间尺度方面，预测这 4种犯罪类型时多使用边长为200 m左右的网格（表2）。由此表明，受预测性警务动态化、精准化防控需求的牵引，以及海量犯罪数据、多源社会经济及建成环境等大数据支持下的精细化研究驱动，犯罪时空预测的分析单元呈现出向微观化发展的趋势，微观尺度的犯罪时空预测已成为当前主流。
表2 2013—2021年对4种犯罪类型进行预测时的常用时空尺度
Tab. 2
| 犯罪类型 | 时间尺度 | 空间尺度 | 常用时空尺度 |
|---|---|---|---|
| 偷窃 | 以天[27-28]为主 | 以边长200 m左右的网格为主[29⇓⇓-32]，最小边长为50 m[33-34]，最大边长为804 m[28] | 天和小网格[28,32] |
| 抢劫 | 以天[17,28]为主 | 以边长200 m左右的网格为主[30,32,35]，最小边长为10 m[36]，最大边长为804 m[28] | 天和小网格[17,28] |
| 入室盗窃 | 以1个月[21,37]为主 | 以边长200 m左右的网格为主[20,30,35]，最大边长为800 m[38] | 1周至半年、小网格[38-39] |
| 盗窃机动车 | 以天[19,32]为主 | 以边长200 m左右的网格为主[20,30,32] | 天和小网格[17,32] |
注：表中小网格指边长在1000 m以内的网格。
4.4 犯罪时空预测的模型与方法
4.4.1 模型概况与常用方法
图5
图5
2013—2021年犯罪时空预测文献中各类方法的数量占比及其年际变化
Fig. 5
The proportion of papers with various methods and its yearly variations in spatial-temporal crime prediction studies from 2013 to 2021
4.4.2 模型精度评价常用指标
在预测模型精度评价方面，相关文献中使用最多的前5个指标分别为：均方根差（Root Mean Squared Error, RMSE, 37篇）、准确率（Accuracy, 35篇）、预测准确率指数（Prediction Accuracy Index, PAI, 24篇）、精准度（Precision, 23篇）和平均绝对值误差（Mean Absolute Error, MAE, 22篇）。其中，RMSE和MAE用于衡量预测值和真实值之间的平均绝对差异，易受异常值的影响，多用于回归任务；Accuracy表示正确预测的单元在所有单元中的占比，受研究区域内案发水平和案件空间集聚程度的影响较大，Precision表示在预测发生案件的空间单元中实际发生案件的单元数量占比，二者多用于分类任务；PAI在犯罪地理学界用得较多，它是采用面积标准化后的准确率指标，由命中率与预测面积占比的比值得到[60]（式（1））。
式中：n指预计会发生犯罪的地区的犯罪数量；N指研究区域内的犯罪数量；a指预计会发生犯罪的区域面积；A指研究区域的面积。
式中：
4.4.3 4类犯罪在不同时空尺度下的优势模型
针对犯罪时空预测研究中的4种常见犯罪类型，基于精度评价相关指标，整理出对4类犯罪在不同时空尺度上进行预测的效果最优模型（表3）。根据筛选出的194篇方法类文献，将时间尺度划分为微观（1个月以下）、中观（1个月至1年）、宏观（1年及以上） 3个层次，将空间尺度也划分为微观（小于既有功能分区的区域）、中观（采用既有功能分区，包括：人口普查区、社区、街区、邮政编码区、警区、交通分析区等）、宏观（区县市及以上区域） 3个层次。
表3 2013—2021年犯罪预测研究中4类犯罪在不同时空尺度下的优势模型
Tab. 3
| 犯罪类型 | 时间尺度 | 空间尺度 | 预测效果 | 模型方法 | 研究区 | |
|---|---|---|---|---|---|---|
| 偷窃 | 微观 | 微观 | RMSE：0.0284[64] | BP神经网络 | 美国芝加哥 | |
| PAI：37.9[65] | 数据驱动的格林函数方法（DDGF）+自激点过程 | 美国芝加哥 | ||||
| 微观 | 中观 | RMSE：1.03[27] | LSTM和时空图卷积网络（ST-GCN）相结合的方法 | 美国芝加哥 | ||
| Precision：0.775[66] | 随机森林 | 菲律宾马尼拉 | ||||
| 抢劫 | 微观 | 微观 | PAI：10.3[69] | 时空协克里金算法 | 中国ZG市XT警区 | |
| 微观 | 中观 | RMSE：0.145[33] | 随机森林 | 菲律宾马尼拉 | ||
| 宏观 | 微观 | Precision：0.775[66] | 随机森林 | 美国达拉斯 | ||
| 宏观 | 中观 | PAI：40[67] | 线性判别分析与K-近邻算法集成模型（LDAKNN） | 中国江西省南昌市主城区 | ||
| RMSE：0.4左右[68] | 基于密度的聚类算法（DBSCAN） | 美国达拉斯 | ||||
| 入室盗窃 | 微观 | 微观 | PAI：15.3[69] | 时空神经网络-门控循环单元（STNN-GRU） | 美国波特兰 | |
| PAI：8 [17] | 加入公平性的霍克斯过程 | 美国印第安纳波利斯 | ||||
| 中观 | 微观 | RMSE：2.264[37] | 综合拉普拉斯近似框架 | 荷兰阿姆斯特丹20个社区 | ||
| Precision：0.24左右[39] | 逻辑回归与多层感知器（MLP）集成模型 | 比利时某城市 | ||||
| PAI：78.5[16] | 旋转网格最大化（RGPM）+随机森林 | 美国波特兰 | ||||
| 宏观 | 中观 | PAI：17.5[69] | 基于密度的聚类（DBSCAN） | 美国达拉斯 | ||
| 盗窃机动车 | 微观 | 微观 | Precision：0.863[70] | 时空神经网络-门控循环单元（STNN-GRU） | 美国波特兰 | |
| PAI：6[17] | 加入公平性的霍克斯过程 | 美国印第安纳波利斯 | ||||
| 中观 | 中观 | PAI：10.3[69] | 基于密度的聚类（DBSCAN） | 美国达拉斯 | ||
注：关于时间和空间维度的宏观、中观、微观尺度划分标准详见4.4.3节正文。
4.5 犯罪时空预测实践评估
当将犯罪时空预测结果用于指导警务实践时，特定警务策略的实施会对未来一段时间内犯罪时空预测模型的结果准确性产生影响。因此，所研究的犯罪时空预测模型能否对警方行动产生持续性的指导价值，有待进行实践评估。由于此类研究需要警方配合并持续投入各种资源，相关研究仍非常少。
目前关于犯罪时空预测实践评估的3个代表性研究表明：因实验区治安条件、实验方案设计、警方配合程度、警务干预策略等多方面的差异，所得出的结论也存在较大差异（表4）。在美国路易斯安那州什里夫波特的实地评估发现，基于多元逻辑回归模型的预测性警务实践与传统警务模式在减少财产犯罪上没有明显区别；而洛杉矶的实验则表明，基于传染型余震序列（ETAS）模型的犯罪时空预测能够显著地减少多类型的犯罪；美国费城的警务实验表明，基于HunchLab软件的预测性警务对财产类犯罪和暴力类犯罪均没有具有统计意义的显著影响，但不同具体警务策略所带来的犯罪数量降幅有明显差异，使用带标记警车对预测高发区域进行巡逻，预期能使财产类犯罪降低31%。
表4 2013—2021年预测性警务效果实地评估的3个代表性研究总结
Tab. 4
| 研究者 | 实验时间 | 实验地点 | 实验设计 | 预测模型/软件 | 犯罪类型 | 预测时空分析单元 | 实验效果评估结论 |
|---|---|---|---|---|---|---|---|
| Hunt等[26] | 2012.06.04—2012.12.21 | 美国路易斯安那州什里夫波特警察局 | 在3个指挥中心下共设计3个控制区和3个试验区： ① 指挥中心1包括2个试验区；② 指挥中心2包括1试验区和1控制区； ③ 指挥中心3包括2控制区 | 多元逻辑回归模型 | 居住区和商业区入室盗窃、盗窃机动车、盗窃机动车内财物 | 时间：每月 空间：犯罪风险概率为“中”（40%~60%）、“高”（60%以上）的122 m×122 m网格 | 预测性警务实践在减少财产犯罪方面的作用与传统警务模式没有明显区别。但研究者并不确定这是由于模型本身缺陷，还是因当地犯罪率过低、各区警务工作差异所致 |
| Mohler等[24] | 2011.11.07— 2012.04.27 2012.03.31— 2012.09.14 2012.05.16— 2013.01.10 | 美国洛杉矶的3个警区 | 由ETAS模型、警局犯罪分析师分别做出犯罪预测，确定最高发的20个或40个网格。随机决定当天采取何种预测结果，并在早晨点名时告知巡逻警察被选用的犯罪预测高发网格。 | 传染型余震序列（ETAS）模型 | 入室盗窃、盗窃机动车、盗窃机动车内财物 | 时间：每天 空间：150 m×150 m空间网格 | 采用ETAS算法，每个预测区中警察巡逻时间每增加1000 min，犯罪数量预期减少1起。相当于在中等巡逻强度（每天每个预测区域31 min）下，可使每个警区每周的平均犯罪数量减少4.3起，降幅为7.4%。采用警局犯罪分析师的预测结果，无统计学显著差异 |
| Ratcliffe等[25] | 2015.06.01— 2015.08.25 2015.11.01— 2016.01.31 | 美国费城 | 将20个警区随机分为4组：1组对照区，采用日常巡逻策略；3组干预区，警务策略分别是：仅告知警员预测高发区域、使用有标记警车巡逻预测高发区域、使用无标记车辆和便衣警员巡逻预测高发区域。 | HunchLab软件 | 财产类犯罪 暴力类犯罪 | 时间：每天8 h；财产类犯罪为 8:00—16:00，暴力类犯罪为 18:00—次日2:00 空间：每个警区内预测犯罪概率最高的3个网格（152 m×152 m） | 财产类犯罪：均无统计显著性影响，但使用标记警车区的犯罪数量平均降幅31%，其他干预区平均降幅不明显。暴力类犯罪：均无统计显著性影响，且各实验区犯罪数量的平均降幅也不明显 |
4.6 犯罪时空预测主要要素的共现关联关系
通过构建犯罪时空预测主要要素的知识关系图谱（图6）可知：当所研究的犯罪类型为较为宽泛的一大类（如：暴力犯罪、财产犯罪）时，预测的时空单元往往是较为宏观的年、周、城市、国家等；当研究的犯罪为较为高发的某一具体类型（如：偷窃、入室盗窃、盗窃机动车等）时，预测的时空单元随之细化为较为微观的天和小尺度格网（以边长为200 m的格网为主）。随着KDE、RTM、决策树、随机森林、LSTM及其他深度学习方法的使用，犯罪时空预测的评价指标不再局限于由犯罪地理学者提出的PAI和PEI，也会采用准确率、精准度、召回率、F1得分等指标。该现象在进行微观时空尺度的犯罪时空预测时较为明显，可能是与较多计算机相关专业的学者开始关注犯罪时空预测有关。
图6
图6
2013—2021年已发表文献中犯罪时空预测主要要素的知识关系图谱
Fig. 6
Knowledge relationship map of main components in spatial-temporal crime prediction studies from 2013 to 2021
5 犯罪时空预测软件及其在各国的应用实践
5.1 代表性犯罪时空预测商业化软件
在犯罪时空预测方面，欧美国家已开发出 Predpol、HunchLab、PreCobs等成熟的商业化软件，并在各国犯罪预测性警务实践中得到了广泛应用。不同软件所针对的犯罪类型、采用的预测模型和算法、时空分析单元、输入数据以及应用推广情况等都存在些许差异。国外一些代表性犯罪时空预测软件的特点和应用情况对比详见表5。当前，我国尚没有开发出具有自主知识产权的商业化犯罪时空预测软件，也未正式引入和推广应用国外的相关软件。
表5 2013—2021年间犯罪时空预测的代表性商业软件特点及应用情况
Tab. 5
| 软件名称 | 预测算法 | 适用犯罪类型 | 输入数据 | 时间步长 | 网格边长 | 应用推广情况 | 应用效果 |
|---|---|---|---|---|---|---|---|
| PredPol （现为Geolitica） | 余震模型与机器学习算法 | 财产类、暴力类犯罪 | 犯罪案件记录 | 一般为8 h | 约150 m | 美国洛杉矶、亚特兰大、圣克鲁斯等近60个城市；英国的伦敦、肯特郡、约克郡等地 | 使用PredPol后，能有效降低犯罪率[3,24,73]。洛杉矶实验后，犯罪数量平均降低7.4%[24] |
| HunchLab (现为ShotSpotter® Missions™) | 机器学习算法 | 财产类、暴力类犯罪 | 辖区边界、犯罪案件记录、地理数据、时态数据 | 一般为 1小时至数小时 | 100~250 m | 美国皮奥里亚、费城、林肯、纽约、新城堡、塔科马、皮尔斯等地 | 费城实验表明，可明显遏制财产类犯罪[25]。芝加哥测试表明，能促进减少暴力犯罪[74] |
| RTMDx | 风险地形建模 | 多种犯罪类型 | 犯罪记录、地理空间数据等 | 一般为 6个月 | 建议为街道平均长度的一半 | 美国堪萨斯城、纽瓦克、大西洋城、纽约、纽黑文、泽西城等地 | 芝加哥：袭击和抢劫有明显减少。纽瓦克：枪支暴力减少35%。科罗拉多斯普林斯：盗窃机动车减少33%[75-76] |
| CAS | 神经网络模型 | 财产类、暴力类犯罪 | 历史犯罪数据、社会经济数据、犯罪机会数据 | 8 h | 125 m | 荷兰全国[77] | 阿姆斯特丹：可准确预测15%的入室盗窃和33%的抢劫，但实践应用效果未知[71] |
| PreCobs | 邻近重复模型 | 仅居住区入室盗窃犯罪 | 历史犯罪数据、住宅类型 | 一般为1 d | 250 m | 德国巴伐利亚州、巴登-符腾堡州等6个州；瑞士的苏黎世、巴塞尔市、阿尔高州等[71] | 德国纽伦堡和慕尼黑：犯罪率下降了14%，严格控制的地区则下降了近30%[78-79]。苏黎世：全市半年的入室盗窃下降了40%[80] |
注：数据根据文献、公司网站及互联网公开资料整理。
5.2 各国预测性警务应用实践
美国是预测性警务的倡导者和推动者。美国洛杉矶警察局（LAPD）于2008年首次提出了预测性警务。2009年，在美国洛杉矶召开了首届预测性警务研讨会；2011年7月美国加利福尼亚州圣克鲁斯警察局开展了首个预测性警务技术测试[3]，标志着预测性警务研究和创新实践的序幕正式拉开。根据“监控图集”（Atlas of Surveillance）网站公开信息统计，截至2021年底，美国有187个州、县或城市警察局通过采购Predpol、HunchLab、RTMDx等商业软件或自行开发等方式（图7（a）），建立了预测性警务应用系统。但在2018年和2019年一批警察机构与Predpol等公司签署服务合同后，近年来采用预测性警务的新增机构数量出现了减缓甚至停滞增长的趋势（图7（b）），这可能与当前美国预测性警务实践正遭受社会各界对其合法性、公平性、种族偏见、公民隐私侵犯等问题的质疑密切相关。
图7
图7
2011—2021年美国实施预测性警务的机构统计
Fig. 7
Statistics of agencies adopting predictive policing in the U.S.A. from 2011 to 2021
英国早期主要使用商业公司（如：Azevea、Palantir、PredPol）开发的犯罪预测软件，但之后由于成本高昂而停止使用[2]。目前，主要采用自行研发的犯罪预测系统，如：由英国伦敦大都会警察局与伦敦大学学院（UCL）共同研发的基于路网的犯罪预测系统。
我国开展过一些预测性警务探索，但尚没有真正实现业务化运行，这可能与国内外警务模式之间存在的较大差异有关。借鉴IBM公司在美国孟菲斯市开发Blue CRUSH（Criminal Reduction Utilizing Statistical History，利用统计历史数据减少犯罪）项目的成功经验，北京市怀柔公安分局于2013年开发了相关应用，实现了警力投量投向的时空引导[81]。江苏省苏州市公安局于2014年开发了一个类似于PredPol的犯罪预测系统，并在2个派出所内开展了试点，据相关报道称取得了不错的应用效果。然而，据了解，由于国内外警务模式的差异，系统运行与我国警务工作机制不相适应，相关系统后续均已停用。在国外，警察预防和打击犯罪的重要手段之一是州、县和城市警察巡逻，而在中国，社区巡逻工作通常是由派出所民警承担，每位民警所负责的警务责任区空间范围相对较小，责任民警对区内整体治安状况和犯罪热点较为熟悉。与责任民警的经验相比，当前大多数犯罪时空预测模型的预测效果并不理想，因此在实践中也就难以得到持续应用。
总之，以美国预测性警务模式为参考，国外很多国家开展了将犯罪时空预测与警务模式相结合的应用实践。我国虽然也较早地同步开展了相关尝试，但由于国内外警务模式之间存在较大差异，并缺乏必要的系统化研究跟进指导，犯罪时空预测结果与国内警务模式的契合程度并不紧密，导致相关实践应用的效果并不理想。
6 犯罪时空预测的伦理问题研究
6.1 预测性警务面临的合法性与伦理问题挑战
近年来，关于预测性警务合法性、公平性、问责制等方面的质疑越来越多[11⇓-13]，给国外相关警务执法部门带来了空前压力。2016年8月31日，美国公民自由联盟（American Civil Liberties Union，ACLU）和16个民权隐私、种族正义和技术组织组成的联盟发表联合声明，指出预测性警务存在种族偏见、缺乏透明度和其他深层次缺陷[82]。2020年5月发生的黑人乔治·弗洛伊德（George Floyd）遭警察“跪杀”事件，进一步加剧了公众对预测性警务应用的不满[83]。2020年6月，美国首个尝试预测性警务计划的圣克鲁斯市成为第一个全面禁止使用犯罪预测技术的城市[84]，其他多个城市也陆续全面或部分停止使用基于人的预测性警务工具[85-86]。此外，2020年5月荷兰法院也判定禁止使用用于预测诈骗者的SyRI(System Risk Indication)系统[87]。
6.2 犯罪时空预测的相关伦理问题
图8
图8
基于地点的预测性警务实践中存在的潜在偏见与伦理问题
Fig. 8
Potential bias and ethical issues in the practice of place-based predictive policing
6.2.1 有偏数据引发的公平性、种族偏见等系列问题
6.2.2 算法偏见和公平性问题
6.2.3 缺乏透明度和问责制
6.3 伦理问题应对
近年来，预测性警务的政策制定者、预测软件开发人员和执法机构也在加强自我反省[2]，一系列有效解决种族偏见、合法性、公平性等问题的建议和成功实施预测性警务的行为准则被提出。例如，Furgson[3]分析了预测性警务的9方面缺陷和应对方法，提出了一个评估分析框架；Moses和Chan[105]分析了成功实施预测性警务四阶段任务的10个理想假设及其所面临的实际障碍和应注意的问题；Mohler等[70]将公平性引入点过程模型中，提出了一种平衡准确性和公平性的惩罚似然法，实现巡逻强度级别与人口统计数据的按比例匹配[70]；HunchLab（www.shotspotter.com/missions/）专门开发了定向巡逻功能，能跟踪警员花费的时间和使用的战术，以避免在特定区域过度巡逻问题。
7 结论与展望
7.1 结论
近年来，预测性警务作为一种数据驱动的前瞻式、主动型警务模式，在欧美国家中被广泛采用，支持其运行的重要技术基础之一是犯罪时空预测。通过对2013年以来世界范围内犯罪时空预测研究与实践应用情况进行梳理，可知：
（1）犯罪时空预测的研究文献快速增加，但主要集中在美、中两国；研究主题日益多元，除仍占主导的时空预测方法研究外，犯罪时空预测警务实践的伦理问题也得到了大量关注。
（2）偷窃、抢劫、入室盗窃、盗窃机动车是被研究最多的四类犯罪，入室盗窃犯罪预测的时间单元以1月为主，而其他三类犯罪预测的时间单元则以1天为主，四类犯罪预测的空间单元均普遍采用边长200米的网格。
（3）针对四类犯罪，在不同时空尺度上进行预测的优势模型有所差别，但总体来看，随机森林、时空神经网络、霍克斯过程模型是预测效果较好的3类模型，在同类型犯罪预测中取得了较高的PAI、RMSE或Precision值。
（4）预测性警务实践在经历快速扩张后进入了减缓甚至停滞阶段，犯罪时空预测中内含的数据偏见、算法偏见等导致警务行动的合法性和伦理问题受到了空前的关注，然而针对犯罪时空预测警务实践的全面、系统、科学评估工作却极其有限，同时与世界各地警务工作机制的结合模式也有待创新。
7.2 展望
总体来看，世界范围内的犯罪时空预测在模型方法和实践应用方面都积累了丰富的经验，已经从早年的技术狂热阶段进入到当前的相对理性反思阶段。展望未来，犯罪时空预测研究的根本目标仍然是：服务犯罪防控实践、为追求社会公平与正义提供科学工具。未来，需要重点加强以下几方面的工作：
（1）加强犯罪时空预测模型的优化与创新。随着计算科学的发展，具有更强预测能力的新模型不断出现，要及时地将其引入到犯罪时空预测之中，也可考虑将其与既有的多种模型进行优化组合。已有研究表明，在进行犯罪时空预测时，除考虑时空依赖性外，将执法行为与犯罪行为之间的交互过程也包含在模型中，可有效地提高模型预测效果[107]。因此，可研究提出能反映更加真实、多元社会-行为过程的犯罪时空预测模型，将犯罪案件记录数据的漏斗效应、各类环境或行为记录数据的有偏性、犯罪者行为与警务行动交互的外部性等多个实际过程考虑在内。针对不同的犯罪类型，加强对不同时空尺度下优势模型的理论与实证研究，为不同情景下的预测性警务应用提供最优模型支持。此外，在通用性模型中加入“本地化”环境因子，也可有效提高针对不同地区的犯罪时空预测模型效果。例如，在中国部分城市的犯罪预测中，可将“城中村”、“群防群治”等因素考虑在内。
（2）提高犯罪时空预测模型的可解释性。当前，基于深度学习框架的犯罪时空预测模型通常具有较高的预测精度，但也具有“黑箱”缺陷，使得在将其迁移至新环境中进行犯罪预测时会具有较大的不确定性。未来，要结合犯罪研究的领域知识，选取有解释力的特征变量以提高模型的可解释性；也可考虑在基于深度学习框架的犯罪时空预测模型中增加注意力机制，通过注意力权重来明确不同特征的相对重要性；还可考虑引入“加性解释模型”，评估不同特征对于模型结果的贡献程度等[108]。
（3）探索符合国内警务实践的“本地化”预测性警务模式。一方面，要加强研究者与警务工作人员的密切合作，避免照搬国外的预测模型算法和警务模式，探索将犯罪时空预测信息与国内警务实践相结合的“本地化”路径或模式。另一方面，适度扩大犯罪时空预测信息的服务面，推进警务工作的社会管理创新。通过信息定期交传、信息智能推送服务等方式，将犯罪时空预测预警信息向社区管理、交通出行、城建规划等相关主管部门或者周边居民主动推送，在服务其他行业的同时，创新“群防群治”工作模式，有针对性地发动群众提前做好犯罪预防及可疑犯罪线索的上报，最大化发挥犯罪时空预测信息效用。
（4）开展预测性警务的科学性、综合性效果评估。采用科学严谨的准实验设计，既要评估预测性警务的各类具体行动所产生的实际犯罪防控效果，也要研究犯罪预测的时空范围、结果形式、与警务人员的信息交传机制等因素对实际犯罪防控效果的影响。在进行实践评估时，除将犯罪防控效果作为测量标准外，还应纳入与实践偏见和伦理相关的评估准则，以更为全面地把握预测性警务实践的综合性效果。
参考文献
Predictive policing and crime control in the United States of America and Europe: trends in a decade of research and the future of predictive policing
[J].There has been a significant focus on predictive policing systems, as law enforcement agents embrace modern technology to forecast criminal activity. Most developed nations have implemented predictive policing, albeit with mixed reactions over its effectiveness. Whilst at its inception, predictive policing involved simple heuristics and algorithms, it has increased in sophistication in the ever-changing technological environment. This paper, which is based on a literature survey, examines predictive policing over the last decade (2010 to 2020). The paper examines how various nations have implemented predictive policing and also documents the impediments to predictive policing. The paper reveals that despite the adoption of predictive software applications such as PredPol, Risk Terrain Modelling, HunchLab, PreMap, PRECOBS, Crime Anticipation System, and Azevea, there are several impediments that have militated against the effectiveness of predictive policing, and these include low predictive accuracy, limited scope of crimes that can be predicted, high cost of predictive policing software, flawed data input, and the biased nature of some predictive software applications. Despite these challenges, the paper reveals that there is consensus by the majority of the researchers on the importance of predictive algorithms on the policing landscape.
Policing predictive policing
[J].
A systematic review on spatial crime forecasting
[J].Predictive policing and crime analytics with a spatiotemporal focus get increasing attention among a variety of scientific communities and are already being implemented as effective policing tools. The goal of this paper is to provide an overview and evaluation of the state of the art in spatial crime forecasting focusing on study design and technical aspects.We follow the PRISMA guidelines for reporting this systematic literature review and we analyse 32 papers from 2000 to 2018 that were selected from 786 papers that entered the screening phase and a total of 193 papers that went through the eligibility phase. The eligibility phase included several criteria that were grouped into: (a) the publication type, (b) relevance to research scope, and (c) study characteristics.The most predominant type of forecasting inference is the hotspots (i.e. binary classification) method. Traditional machine learning methods were mostly used, but also kernel density estimation based approaches, and less frequently point process and deep learning approaches. The top measures of evaluation performance are the Prediction Accuracy, followed by the Prediction Accuracy Index, and the F1-Score. Finally, the most common validation approach was the train-test split while other approaches include the cross-validation, the leave one out, and the rolling horizon.Current studies often lack a clear reporting of study experiments, feature engineering procedures, and are using inconsistent terminology to address similar problems.There is a remarkable growth in spatial crime forecasting studies as a result of interdisciplinary technical work done by scholars of various backgrounds. These studies address the societal need to understand and combat crime as well as the law enforcement interest in almost real-time prediction.Although we identified several opportunities and strengths there are also some weaknesses and threats for which we provide suggestions. Future studies should not neglect the juxtaposition of (existing) algorithms, of which the number is constantly increasing (we enlisted 66). To allow comparison and reproducibility of studies we outline the need for a protocol or standardization of spatial forecasting approaches and suggest the reporting of a study's key data items.© The Author(s) 2020.
Spatio-temporal crime hotspot detection and prediction: a systematic literature review
[J].
犯罪时空预测方法研究综述与展望
[J].
DOI:10.12082/dqxxkx.2021.200247
[本文引用: 3]
犯罪时空预测作为预测警务的核心支撑技术,自2000年左右至今得到了快速的发展。本文介绍了犯罪时空预测的实践背景和理论基础,将犯罪时空预测解构为利用历史案件的时空位置、时空环境和个体行为等要素,结合相应的算法模型预测未来案件时空分布的过程。然后,从输入要素的视角对当前的犯罪时空预测方法进行了总结和归纳,将其划分为基于案件时空位置信息的犯罪时空预测、基于时空环境要素的犯罪时空预测,以及融合行为轨迹和时空环境要素的犯罪时空预测3种类型,详细总结了不同类型犯罪时空预测的方法原理,并从适应场景和预测效果等方面对不同的方法模型进行了比较。最后,结合当前的大数据技术发展趋势,对未来的犯罪时空预测进行了展望。本文认为犯罪时空预测未来需要从数据角度重点解决输入数据的体系融合、粒度细化和新型数据融合等问题,从模型优化角度应着重提高多源异构数据融合能力,平衡模型的可解释性与预测效果。
Overview and prospect for spatial-temporal prediction of crime
[J].
“The computer said so”: On the ethics, effectiveness, and cultural techniques of predictive policing
[J].
The politics and biases of the Crime Anticipation System of the Dutch police
[C]//
Ethics, artificial intelligence, and risk assessment
[J].
Predictive policing and civilian oversight: what will it take to get it right?
[J].
AI ethics in predictive policing: from models of threat to an ethics of care
[J].
The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration
[J].
Privacy preserving, crowd sourced crime Hawkes processes
[C]//
Rotational grid, PAI-maximizing crime forecasts
[J].
Area-specific crime prediction models
[C]//
Automatic optimization of localized kernel density estimation for hotspot policing
[C]//
Predicting crime with routine activity patterns inferred from social media
[C]//
A theory-driven algorithm for real-time crime hot spot forecasting
[J].Real-time crime hot spot forecasting presents challenges to policing. There is a high volume of hot spot misclassifications and a lack of theoretical support for forecasting algorithms, especially in disciplines outside the fields of criminology and criminal justice. Transparency is particularly important as most hot spot forecasting models do not provide their underlying mechanisms. To address these challenges, we operationalize two different theories in our algorithm to forecast crime hot spots over Portland and Cincinnati. First, we use a population heterogeneity framework to find places that are consistent hot spots. Second, we use a state dependence model of the number of crimes in the time periods prior to the predicted month. This algorithm is implemented in Excel, making it extremely simple to apply and completely transparent. Our forecasting models show high accuracy and high efficiency in hot spot forecasting in both Portland and Cincinnati context. We suggest previously developed hot spot forecasting models need to be reconsidered.
Mastering the game of Go without human knowledge
[J].
Mastering the game of Go with deep neural networks and tree search
[J].
Randomized controlled field trials of predictive policing
[J].
The Philadelphia predictive policing experiment
[J].
Risk prediction of theft crimes in urban communities: An integrated model of LSTM and ST-GCN
[J].
Spatio-temporal prediction of Baltimore crime events using CLSTM neural networks
[J].
Crime geographical displacement: testing its potential contribution to crime prediction
[J].Crime geographical displacement has been examined in many Western countries. However, little is known about its existence, distribution, and potential predictive ability in large cities in China. Compared to the existing research, this study contributes to the current research in three ways. (1) It provides confirmation that crime geographical displacement exists in relation to burglaries that occur in a large Chinese city. (2) A crime geographical displacement detector is proposed, where significant displacements are statistically detected and geographically displayed. Interestingly, most of the displacements are not very far from one another. These findings confirm the inferences in the existing literature. (3) Based on the quantitative results detected by the crime geographical displacement detector, a crime prediction method involving crime geographical displacement patterns could improve the accuracy of the empirical crime prediction method by 7.25% and 3.1 in the capture rate and prediction accuracy index (PAI), respectively. Our current study verifies the feasibility of crime displacement for crime prediction. The feasibility of the crime geographical displacement detector and results should be verified in additional areas.
Predicting crime using Twitter and kernel density estimation
[J].
Predicting time and location of future crimes with recommendation methods
[J].
Spatial crime distribution and prediction for sporting events using social media
[J].
Crime prediction with historical crime and movement data of potential offenders using a spatio-temporal cokriging method
[J].Crime prediction using machine learning and data fusion assimilation has become a hot topic. Most of the models rely on historical crime data and related environment variables. The activity of potential offenders affects the crime patterns, but the data with fine resolution have not been applied in the crime prediction. The goal of this study is to test the effect of the activity of potential offenders in the crime prediction by combining this data in the prediction models and assessing the prediction accuracies. This study uses the movement data of past offenders collected in routine police stop-and-question operations to infer the movement of future offenders. The offender movement data compensates historical crime data in a Spatio-Temporal Cokriging (ST-Cokriging) model for crime prediction. The models are implemented for weekly, biweekly, and quad-weekly prediction in the XT police district of ZG city, China. Results with the incorporation of the offender movement data are consistently better than those without it. The improvement is most pronounced for the weekly model, followed by the biweekly model, and the quad-weekly model. In sum, the addition of offender movement data enhances crime prediction, especially for short periods.
The effect of spatiotemporal resolution on predictive policing model performance
[J].
The use of predictive analysis in spatiotemporal crime forecasting: building and testing a model in an urban context
[J].
基于临近相似性考虑的犯罪热点密度图预测准确性比较——以DP半岛街头抢劫犯罪为例
[J].
DOI:10.13249/j.cnki.sgs.2016.01.007
[本文引用: 2]
在无时空考虑的密度估计算法基础上,分别加入了案件点之间的时间临近相似性、空间临近相似性和时空临近相似性的考虑,利用DP半岛2006~2007年的街头抢劫犯罪数据为基础计算无时空临近相似性、时间临近相似性、空间临近相似性和时空临近相似性4种不同算法所得到的犯罪热点图,并以之预测2008年的街头抢劫。通过Natural breaks（Jenks）分级方法和等比例面积选取两种方式来划定热点区域进行预测并进行PAI指数得分比较,结果表明时空临近相似性的密度估计算方法在犯罪预测的优势比较显著。
The comparison of predictive accuracy of crime hotspot density maps with the consideration of the near similarity: A case study of robberies at DP peninsula
[J].
Forecasting spatio-temporal variation in residential burglary with the integrated Laplace approximation framework: Effects of crime generators, street networks, and prior crimes
[J].We investigate the spatio-temporal variation of monthly residential burglary frequencies across neighborhoods as a function of crime generators, street network features and temporally and spatially lagged burglary frequencies. In addition, we evaluate the performance of the model as a forecasting tool.
Hierarchical spatio-temporal pattern discovery and predictive modeling
[J].
Comparison of near-repeat, machine learning and risk terrain modeling for making spatiotemporal predictions of crime
[J].
Crime event prediction with dynamic features
[J].
基于犯罪空间分异和建成环境的公共场所侵财犯罪热点预测
[J].
DOI:10.12082/dqxxkx.2019.190358
[本文引用: 1]
机器学习是当前犯罪热点预测的主流方法,随机森林算法因需要的数据量较小、有较好的预测能力和预测精确度、且有较高的可理解度,更是被广泛应用,代表地理环境和建成环境的多源数据也被广泛用于模型改进的尝试实践中,但这些实践都只考虑研究区整体的预测精度变化情况,并未区分不同区域犯罪热点预测结果的差异及其原因。因此,本文以公共场所侵财犯罪为例,根据历史犯罪分布情况及过往犯罪热点分布规律,将研究区分为稳定高发热点网格、较高发热点网格、偶发热点网格及非热点网格这4类,并依据社会失序理论、日常活动理论和犯罪模式理论,选取城中村范围、路网密度及POI（餐饮、娱乐、商场3类设施）密度这3个具有代表性的协变量加入到随机森林预测模型中,探讨预测结果精度的变化情况。根据2017年26个双周的犯罪热点预测实验的预测结果,得到以下结论：加入协变量后,研究区整体、稳定高发热点网格及较高发热点网格的预测精度都有不同程度的提高,分区模型的精度显著高于整体模型的精度,说明考虑空间分异对提高模型精度起重要作用。
Hotspot prediction of public property crime based on spatial differentiation of crime and built environment
[J].
基于随机森林和时空核密度方法的不同周期犯罪热点预测对比
[J].
DOI:10.18306/dlkxjz.2018.06.003
[本文引用: 1]
犯罪预测对于制定警务策略、实施犯罪防控具有重要意义。机器学习和核密度是2类主流犯罪热点预测方法,然而目前还鲜有研究对这2类方法在不同时间周期下的犯罪预测效果进行系统比较,本文试图对此进行补充。本文以2013-2016年5月的公共盗窃犯罪历史数据作为输入,分别对比了在接下来2周、1个月、2个月、3个月4个不同时间周期随机森林方法与基于时空邻近性的核密度方法的犯罪热点预测效果,结果发现：在各时间周期上,随机森林分类热点预测方法的面积和案件量命中率均比时空核密度方法准确性高;并且2种方法均能有效地识别犯罪热点中的高发区域,其中在较小范围较短时间内随机森林识别热点中的高发区效率更高,而在较大范围较长时间周期上时空核密度方法识别高发区更优。
Comparison of random forest algorithm and space-time kernel density mapping for crime hotspot prediction
[J].Crime prediction is of great significance for the formulation of police tactics and the implementation of crime prevention and control in different time periods. Machine learning and density mapping are two common approaches for crime hotspot prediction. However, there exists few published work that systematically compares the predicted results of these two approaches. This study aimed to fill the gap. With crime patterns uncovered from 2013 to May 2016, we predicted hot-spot distribution of theft crimes in the period of first two weeks of June, July, and August in 2016 by random forest algorithm and traditional space-time kernel density method and compared the two sets of predictions. The research area was divided into grid cells of 50 m×50 m. Each cell was predicted as either hot-spot or non-hot-spot area in the next predicting period. Then we overlaid the forecast results and location of real cases to evaluate the accuracy of the two methods. The results show that both the hit rate of area and cases of the random forest classification hot-spot prediction method are higher than that of the space-time kernel density within different periods. Both methods can effectively identify high-crime areas of crime hot spots in prediction. In a relatively short period of time and small area, the random forest classification hotspot prediction method is more effective than the space-time kernel density method. However, in a relatively long term and large area, the space-time kernel density crime risk estimation method yields better result in identifying high crime areas.
Applying crime prediction techniques to Japan: a comparison between risk terrain modeling and other methods
[J].
Using vulnerability and exposure to improve robbery prediction and target area selection
[J].
风险地形建模在犯罪风险评估中的应用
[J].
The application of risk terrain model in risk assessment of crime
[J].
基于风险地形建模的毒品犯罪风险评估和警务预测
[J].
DOI:10.18306/dlkxjz.2018.08.012
[本文引用: 1]
犯罪具有明显的时空特征,研究犯罪问题离不开时间和空间维度分析,以及产生犯罪的社会、地理、生态、环境等因素。风险地形建模是美国学者研发的空间风险评估和警务预测技术,已在全球六大洲45个国家和美国35个州得到了独立测试和验证,被广泛应用于警务预测、国土安全、交通事故、公共医疗、儿童虐待、环境污染、城市发展等多个领域。在毒品、纵火、爆炸、强奸、抢劫、盗窃等犯罪研究领域更是取得了显著成果。本文运用犯罪热点分析和风险地形建模,以长三角地区N市毒品犯罪为研究对象,对该市2015年毒品犯罪的危险因子、空间盲区、风险地形进行分析,探索毒品犯罪的生成机理和演化规律,并对2016年毒品犯罪进行预测。研究结果表明,N市毒品犯罪呈现明显的犯罪热点和冷点;出租屋、酒店、车站、ATM机、停车场、娱乐场所、城市快速路、网吧是N市毒品犯罪的风险性因素。风险地形建模能较好地预测毒品犯罪。公安机关禁毒部门应据此进行严密管控,逐步限制、消除犯罪产生地、犯罪吸引地、犯罪促进地的生存土壤和条件。
Drug-related crime risk assessment and predictive policing based on risk terrain modeling
[J].Crime is the product of a certain time and space. Research on crime cannot be separated from temporal and spatial analyses, as well as social, geographical, ecological, environmental, and other factors that generate crime. Risk terrain modeling technology was developed by American scholars for spatial risk assessment and predictive policing. It has been independently proven and tested in over 45 countries across six continents around the world and 35 states in the United States. It has been widely used in many fields such as predictive policing, homeland security, traffic accidents, public health, child abuse, environmental pollution, and urban development. It has achieved remarkable results in the crime research area of drug, arson, explosion, rape, robbery, and theft. This study adopted crime hotspot analysis and risk terrain modeling to analyze the risk factors, spatial blind spots, and risk terrain of narcotics crimes in 2015 in N City of the Yangtze River Delta region, explored the mechanism and evolution of drug crimes, and made a prediction on N City 2016 drug crime trend. The results show that N City drug crime presents obvious crime hotspots and crime cold spots. Rental housing, hotels, railway stations, banks, parking lots, entertainment venues, urban expressways, and Internet cafes are drug risk factors in the city. Risk terrain modeling is effective in predicting drug crimes. The narcotics departments of public security organs should put more police and energy to gradually limit and eliminate the hotspots that generate, attract, and promote crime.
A spatio-temporal kernel density estimation framework for predictive crime hotspot mapping and evaluation
[J].
Proactive spatiotemporal resource allocation and predictive visual analytics for community policing and law enforcement
[J].
ESCORT: Fine-grained urban crime risk inference leveraging heterogeneous open data
[J].
Crime hotspot prediction based on dynamic spatial analysis
[J].
Spatio-temporal crime predictions by leveraging artificial intelligence for citizens security in smart cities
[J].
Spatio-temporal crime predictions in smart cities: A data-driven approach and experiments
[J].
Big data analytics and mining for effective visualization and trends forecasting of crime data
[J].
基于长短期记忆模型的入室盗窃犯罪预测研究
[J].
Prediction of burglary crime based on LSTM
[J].
Cyber-attack method and perpetrator prediction using machine learning algorithms
[J].Cyber-attacks have become one of the biggest problems of the world. They cause serious financial damages to countries and people every day. The increase in cyber-attacks also brings along cyber-crime. The key factors in the fight against crime and criminals are identifying the perpetrators of cyber-crime and understanding the methods of attack. Detecting and avoiding cyber-attacks are difficult tasks. However, researchers have recently been solving these problems by developing security models and making predictions through artificial intelligence methods. A high number of methods of crime prediction are available in the literature. On the other hand, they suffer from a deficiency in predicting cyber-crime and cyber-attack methods. This problem can be tackled by identifying an attack and the perpetrator of such attack, using actual data. The data include the type of crime, gender of perpetrator, damage and methods of attack. The data can be acquired from the applications of the persons who were exposed to cyber-attacks to the forensic units. In this paper, we analyze cyber-crimes in two different models with machine-learning methods and predict the effect of the defined features on the detection of the cyber-attack method and the perpetrator. We used eight machine-learning methods in our approach and concluded that their accuracy ratios were close. The Support Vector Machine Linear was found out to be the most successful in the cyber-attack method, with an accuracy rate of 95.02%. In the first model, we could predict the types of attacks that the victims were likely to be exposed to with a high accuracy. The Logistic Regression was the leading method in detecting attackers with an accuracy rate of 65.42%. In the second model, we predicted whether the perpetrators could be identified by comparing their characteristics. Our results have revealed that the probability of cyber-attack decreases as the education and income level of victim increases. We believe that cyber-crime units will use the proposed model. It will also facilitate the detection of cyber-attacks and make the fight against these attacks easier and more effective.
A comparative evaluation of spatio temporal deep learning techniques for crime prediction
[C]//
Addressing crime situation forecasting task with temporal graph convolutional neural network approach
[C]//
GeST: a grid embedding based spatio-temporal correlation model for crime prediction
[C]//
基于机器学习的犯罪预测综述
[J].
An overview on crime prediction through machine learning
[J].
The utility of hotspot mapping for predicting spatial patterns of crime
[J].
The "hottest" part of a hotspot: comments on “the utility of hotspot mapping for predicting spatial patterns of crime”
[J].
A research note on the methodological and theoretical considerations for assessing crime forecasting accuracy with the predictive accuracy index
[J].
Application of grid management in spatio-temporal prediction of crime
[C]//
Crime prediction by data-driven Green’s function method
[J].
Geospatial-temporal analysis and classification of criminal data in Manila
[C]//
Mapping the risk terrain for crime using machine learning
[J].
Mixed spatial-temporal characteristics based crime hot spots prediction
[C]//
Redrawing hot spots of crime in Dallas, Texas
[J].In this work we evaluate the predictive capability of identifying long term, micro place hot spots in Dallas, Texas. We create hot spots using a clustering algorithm, using law enforcement cost of responding to crime estimates as weights. Relative to the much larger current hot spot areas defined by the Dallas Police Department, our identified hot spots are much smaller (under 3 square miles), and capture crime cost at a higher density. We also show that the clustering algorithm captures a wide array of hot spot types; some one or two addresses, some street segments, and others an agglomeration of larger areas. This suggests identifying hot spots based on a specific unit of aggregation (e.g. addresses, street segments), may be less efficient than using a clustering technique in practice.
Crime hot spot forecasting: A recurrent model with spatial and temporal information
[C]//
Predictive policing as a new tool for law enforcement? recent developments and challenges
[J].
Algorithmic prediction in policing: assumptions, evaluation, and accountability
[J].
Crime in context: Utilizing risk terrain modeling and conjunctive analysis of case configurations to explore the dynamics of criminogenic behavior settings
[J].Risk terrain modeling (RTM) is a geospatial crime analysis tool designed to diagnose environmental risk factors for crime and identify the places where their spatial influence is collocated to produce vulnerability for illegal behavior. However, the collocation of certain risk factors’ spatial influences may result in more crimes than the collocation of a different set of risk factors’ spatial influences. Absent from existing RTM outputs and methods is a straightforward method to compare these relative interactions and their effects on crime. However, as a multivariate method for the analysis of discrete categorical data, conjunctive analysis of case configurations (CACC) can enable exploration of the interrelationships between risk factors’ spatial influences and their varying effects on crime occurrence. In this study, we incorporate RTM outputs into a CACC to explore the dynamics among certain risk factors’ spatial influences and how they create unique environmental contexts, or behavior settings, for crime at microlevel places. We find that most crime takes place within a few unique behavior settings that cover a small geographic area and, further, that some behavior settings were more influential on crime than others. Moreover, we identified particular environmental risk factors that aggravate the influence of other risk factors. We suggest that by focusing on these microlevel environmental crime contexts, police can more efficiently target their resources and further enhance place-based approaches to policing that fundamentally address environmental features that produce ideal opportunities for crime.
Vulnerability and exposure to crime: applying risk terrain modeling to the study of assault in Chicago
[J].
Review of state of the art: Predictive policing
[R].
犯罪预测时空定位信息管理系统的构建
[J].
Construction of information management system for predicting crime location in time and space
[J].
Mathematicians urge colleagues to boycott police work in wake of killings
[J].
Technology can't predict crime, it can only weaponize proximity to policing
[R].
Predictive policing: the risks associated with risk assessment
[J].Focusing on the Dutch tools SyRI and CAS, this paper describes predictive policing against the background of the broader development toward a pre-crime society, the accompanying culture of control and the new penal logic it gives rise to. It will explain the risks associated with the risk assessments predictive policing tools provide and end with the recommendation to use predictive policing not only for police deployment, but also to target problem-oriented responses to crime to the right persons and places.
The relationship between crime reporting and police: implications for the use of uniform crime reports
[J].
To predict and serve?
[J].
Machine learning models for predicting crime hotspots in Medellin city
[C]//
Ability of crime, demographic and business data to forecast areas of increased violence
[J].Identifying geographic areas and time periods of increased violence is of considerable importance in prevention planning. This study compared the performance of multiple data sources to prospectively forecast areas of increased interpersonal violence. We used 2011-2014 data from a large metropolitan county on interpersonal violence (homicide, assault, rape and robbery) and forecasted violence at the level of census block-groups and over a one-month moving time window. Inputs to a Random Forest model included historical crime records from the police department, demographic data from the US Census Bureau, and administrative data on licensed businesses. Among 279 block groups, a model utilizing all data sources was found to prospectively improve the identification of the top 5% most violent block-group months (positive predictive value = 52.1%; negative predictive value = 97.5%; sensitivity = 43.4%; specificity = 98.2%). Predictive modelling with simple inputs can help communities more efficiently focus violence prevention resources geographically.
Mitigating vulnerabilities through forecasting and crime trend analysis
[C]//
Predicting spatial crime occurrences through an efficient ensemble-learning model
[J].While the use of crime data has been widely advocated in the literature, its availability is often limited to large urban cities and isolated databases that tend not to allow for spatial comparisons. This paper presents an efficient machine learning framework capable of predicting spatial crime occurrences, without using past crime as a predictor, and at a relatively high resolution: the U.S. Census Block Group level. The proposed framework is based on an in-depth multidisciplinary literature review allowing the selection of 188 best-fit crime predictors from socio-economic, demographic, spatial, and environmental data. Such data are published periodically for the entire United States. The selection of the appropriate predictive model was made through a comparative study of different machine learning families of algorithms, including generalized linear models, deep learning, and ensemble learning. The gradient boosting model was found to yield the most accurate predictions for violent crimes, property crimes, motor vehicle thefts, vandalism, and the total count of crimes. Extensive experiments on real-world datasets of crimes reported in 11 U.S. cities demonstrated that the proposed framework achieves an accuracy of 73% and 77% when predicting property crimes and violent crimes, respectively.
Public decision support for low population density areas: An imbalance-aware hyper-ensemble for spatio-temporal crime prediction
[J].
Epistemologies of predictive policing: Mathematical social science, social physics and machine learning
[J].
The ethical dangers and merits of predictive policing
[J].Predictive policing is an emerging law enforcement technique that uses data and statistical analysis to aid in the identification of criminal activity. Its intention is to proactively reduce crime by providing police forces with likely areas of high risk variables. While this is a noble pursuit, every new tool must be accompanied by the ethical considerations of its potential consequences. Predictive policing is still in its infancy, borne from crime analysis and big data; however, the Western criminal justice system in the traditional sense is a reactive institution with a diverse history. The use of predictive policing presents a new challenge for law enforcement in that it allows for a divergence from the distinct reality of modern policing. Using the United States as an example of the dangers and flaws of predictive policing as a discretionary tool used to justify questionable processes and biases, this paper will analyze the potential opportunity that predictive policing and new holistic forms of law enforcement and community safety initiatives can use in partnerships with communities and policy makers.
Reform predictive policing
[J].
Fairness, accountability and transparency: Notes on algorithmic decision-making in criminal justice
[J].Over the last few years, legal scholars, policy-makers,\n activists and others have generated a vast and\n rapidly expanding literature concerning the ethical\n ramifications of using artificial intelligence,\n machine learning, big data and predictive software\n in criminal justice contexts. These concerns can be\n clustered under the headings of fairness,\n accountability and transparency. First, can we trust\n technology to be fair, especially given that the\n data on which the technology is based are biased in\n various ways? Second, whom can we blame if the\n technology goes wrong, as it inevitably will on\n occasion? Finally, does it matter if we do not know\n how an algorithm works or, relatedly, cannot\n understand how it reached its decision? I argue\n that, while these are serious concerns, they are not\n irresolvable. More importantly, the very same\n concerns of fairness, accountability and\n transparency apply, with even greater urgency, to\n existing modes of decision-making in criminal\n justice. The question, hence, is comparative: can\n algorithmic modes of decision-making improve upon\n the status quo in criminal justice? There is\n unlikely to be a categorical answer to this\n question, although there are some reasons for\n cautious optimism.
Towards a crime hotspot detection framework for patrol planning
[C]//
On the best-performed time window size for homicide count forecasting
[C]//
Pitfalls of predictive policing: An ethical analysis
[J].
Is big data challenging criminology?
[J].The advent of ‘Big Data’ and machine learning algorithms is predicted to transform how we work and think. Specifically, it is said that the capacity of Big Data analytics to move from sampling to census, its ability to deal with messy data and the demonstrated utility of moving from causality to correlation have fundamentally changed the practice of social sciences. Some have even predicted the end of theory—where the question why is replaced by what—and an enduring challenge to disciplinary expertise. This article critically reviews the available literature against such claims and draws on the example of predictive policing to discuss the likely impact of Big Data analytics on criminological research and policy.
Conflicts and congruencies between predictive policing and the patrol officer's craft
[J].
Event-level prediction of urban crime reveals a signature of enforcement bias in US cities
[J].
Interpretable machine learning models for crime prediction
[J].
