A Police Incident Prediction Model Based on Multi-Scale Spatio-Temporal Graph Fusion

doi:10.11871/jfdc.issn.2096-742X.2026.02.012

Abstract

Abstract:

[Objective] Accurate spatio-temporal prediction of police incidents is crucial for optimizing resource allocation and enhancing crime prevention capabilities. [Application Background] Current approaches often suffer from limitations such as single-scale modeling and inadequate adaptation to multi-scale characteristics. To address these issues, this study proposes a novel police incident prediction model based on a Multi-Scale Spatio-Temporal Graph Fusion Network (MS-STGFN). [Methods] The model constructs a dual-scale spatial correlation system incorporating both administrative divisions and road networks, enabling multi-level feature extraction and dynamic fusion of spatio-temporal heterogeneities in incident distributions. Specifically, an adaptive adjacency matrix generation mechanism is designed to integrate cross-scale spatial correlations between macro administrative layouts and micro road network topologies. Subsequently, a hierarchical graph convolutional network incorporating a multi-head attention mechanism facilitates dynamic interaction and collaborative learning of incident features across different scales. Finally, the extracted spatio-temporal features are fused with multi-source urban data to predict the occurrence and distribution of incidents. [Results] Experimental studies conducted in Chaoyang District, Beijing, demonstrate that the proposed model effectively overcomes the limitations of traditional single-scale models in fully capturing spatial heterogeneity. It achieves an F1-score of 0.7729 and an RMSE of 0.4008, outperforming baseline models with a 1.3% improvement in F1-score and a 4.8% reduction in RMSE. [Conclusions] This study provides a new technical pathway for multi-source spatio-temporal data fusion and offers a significant decision-support tool for the development of smart policing and precise crime prevention strategies.

Key words: 110 police incidents, spatio-temporal prediction, multi-scale spatio-temporal modeling, graph attention network, road network modeling, multi-modal feature fusion

JU Zizheng, CHEN Peng, SUI Jinguang, ZHU Longsheng. A Police Incident Prediction Model Based on Multi-Scale Spatio-Temporal Graph Fusion[J]. Frontiers of Data and Computing, 2026, 8(2): 154-170, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.02.012.

Figures/Tables 14

Fig.1

Fig.2

Fig.3

Fig.4

Fig.5

Fig.6

Fig.7

Table 1

Table 2

Table 3

Table 4

Fig.8

Fig.9

Table 5

References 31

[1]	王世卿. 110接处警的历史、现状、问题及规范化研究[J]. 中国人民公安大学学报(社会科学版), 2016, 32(6): 98-104.
[2]	师容. “公安110”供给模式创新研究[J]. 中国人民公安大学学报(社会科学版), 2018, 34(3): 150-156.
[3]	陈伟, 王珩, 尉志凌. 公安机关视野下的社会治安管理创新研究[M]. 甘肃人民出版社, 2013: 50-65
[4]	陈鹏, 赵鹏凯, 瞿珂. 基于时间序列模型的110警情数据预测研究[J]. 信息系统工程, 2015, 000(009): 128-131.
[5]	MOHLER G O, SHORT M B, BRANTINGHAM P J, et al. Self-exciting point process modeling of crime[J]. Journal of the American Statistical Association, 2011, 106(493): 100-108. doi: 10.1198/jasa.2011.ap09546
[6]	ZHANG X, LIU L, XIAO L, et al. Comparison of machine learning algorithms for predicting crime hotspots[J]. IEEE Access, 2020, 8: 181302-181310. doi: 10.1109/Access.6287639
[7]	柳林, 纪佳楷, 宋广文, 等. 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]. 地球信息科学学报, 2019, 21(11): 1655-1668. doi: 10.12082/dqxxkx.2019.190358
[8]	ALWEE R, SHAMSUDDIN S M, SALLEHUDDIN R. Modeling of violent crime rates with economic indicators using hybridization of grey relational analysis and support vector regression[J]. Jurnal Teknologi, 2015, 76(1): 229-238.
[9]	ALDOSSA RI B S, ALQAHTANI F M, ALSHAHRANI N S, et al. A comparative study of decision tree and naive bayes machine learning model for crime category prediction in Chicago[C]// Proceedings of 2020 6th International Conference on Computing and Data Engineering, 2020: 34-38.
[10]	HUANG C, ZHANG C, ZHAO J, et al. Mist: A multiview and multimodal spatial-temporal learning framework for citywide abnormal event forecasting[C]// The World Wide Web Conference, 2019: 717-728.
[11]	SUN J, YUE M, LIN Z, et al. Crimeforecaster: Crime prediction by exploiting the geographical neighborhoods’ spatiotemporal dependencies[C]// Machine Learning and Knowledge Discovery in Databases. Applied Data Science and Demo Track: European Conference, ECML PKDD 2020, Ghent, Belgium, September 14-18, 2020, Proceedings, Part V. Springer International Publishing, 2021: 52-67.
[12]	卢业成, 陈鹏, 江欢, 等. 犯罪时间序列的混沌特征分析与短期预测[J]. 科学技术与工程, 2023, 23(11): 4693-4701.
[13]	MESKELA T E, AFEWORK Y K, AYELE N A, et al. Designing time series crime prediction model using long short-term memory recurrent neural network[J]. International Journal of Recent Technology and Engineering (IJRTE), 2020, 9(4): 402-405.
[14]	王娟, 樊畅, 管雨翔. 基于CNN-LSTM-Attention模型的盗窃犯罪分析与预测[J]. 大数据时代, 2024, (10): 27-40.
[15]	MUTHAMIZHARASAN M, PONNUSAMY R. Forecasting crime event rate with a CNN-LSTM model[M]//Innovative Data Communication Technologies and Application: Proceedings of ICIDCA 2021. Singapore: Springer Nature Singapore, 2022: 461-470.
[16]	ZHANG H, ZHANG J, WANG Z, et al. An adaptive spatial resolution method based on the ST-ResNet model for hourly property crime prediction[J]. ISPRS International Journal of Geo-Information, 2021, 10(5): 314. doi: 10.3390/ijgi10050314
[17]	WANG C, LIN Z, YANG X, et al. Hagen: Homophily-aware graph convolutional recurrent network for crime forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence, 2022, 36(4): 4193-4200.
[18]	WEI C, LI W, ZHANG Y, et al. MVST: A Multi-View Spatial-Temporal Model for Fine-Grained Crime Prediction[J]. ACM Transactions on Intelligent Systems and Technology, 2025, 16(2): 1-22.
[19]	DONG Z, MATEU J, XIE Y. Spatio-temporal-network point processes for modeling crime events with landmarks[J]. arXiv preprint arXiv:2409.10882, 2024.
[20]	贺日兴, 唐宗棣, 姜超, 等. 顾及道路权重的图卷积犯罪时空预测模型[J]. 地球信息科学学报, 2023, 25(10): 1986-1999. doi: 10.12082/dqxxkx.2023.230299
[21]	SUI J, CHEN P, GU H. Deep Spatio-Temporal Graph Attention Network for Street-Level 110 Call Incident Prediction[J]. Applied Sciences, 2024, 14(20): 9334. doi: 10.3390/app14209334
[22]	林逸航, 郑坤, 夏书豪, 等. 融合区域空间相似性特征与事件时空特征的犯罪预测模型[J/OL]. 武汉大学学报(信息科学版), 1-19[2025-08-19]. https://doi.org/10.13203/j.whugis20230395.
[23]	SAK H, SENIOR A, BEAUFAYS F. Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition[J]. arXiv preprint arXiv:1402.1128, 2014.
[24]	RATCLIFFE J H. Geocoding crime and a first estimate of a minimum acceptable hit rate[J]. International Journal of Geographical Information Science, 2004, 18(1): 61-72. doi: 10.1080/13658810310001596076
[25]	BREIMAN L. Random forests[J]. Machine learning, 2001, 45: 5-32. doi: 10.1023/A:1010933404324
[26]	Wang D, Cui P, Zhu W. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining[J]. Structural deep network embedding, 2016: 1225-1234.
[27]	GUO H, TANG R, YE Y, et al. DeepFM: A factorization-machine based neural network for CTR prediction[J]. arXiv preprint arXiv:1703.04247, 2017.
[28]	XIAO J, YE H, HE X, et al. Attentional factorization machines: Learning the weight of feature interactions via attention networks[J]. arXiv preprint arXiv:1708.04617, 2017.
[29]	WANG R, FU B, FU G, et al. Deep & cross network for ad click predictions[M]//Proceedings of the ADKDD'17, 2017:1-7.
[30]	HE X, CHUA T S. Neural factorization machines for sparse predictive analytics[C]// Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2017: 355-364.
[31]	YU L, DU B, HU X, et al. Deep spatio-temporal graph convolutional network for traffic accident prediction[J]. Neurocomputing, 2021, 423: 135-147. doi: 10.1016/j.neucom.2020.09.043

特征类型	变量名	数据类型	取值范围/示例	描述说明
时间特征	month	整数	[1-12]	月份
	day	整数	[1-31]	日
	hour	整数	[0-24]	小时（24小时制）
	weekday	整数	[1-7]	周
气象特征	wind	浮点数	[0.1, …] (m/s)	风速
	temp	浮点数	[-4.7, …] (℃)	温度
	precipitation	整数	{0,1}	降水（0：无，1：有）
	pressure	浮点数	[1022.4, …]	气压
	humidity	整数	[17, …]	湿度
社会特征	is_holiday	整数	{0,1}	节假日标记（0：无，1：有）

特征类别	区域类型	子类别	特征指标
空间特征	道路网络	道路属性	单向通行、层级、桥梁、隧道、道路类型、长度
		人口分布	人口数量/密度
		POI分布	餐饮、景点、公共、企业、购物等13类POI数量/密度
	行政区划	行政区划属性	行政区划类型、面积、区划边界长度、邻接行政区划数量
		人口分布	人口数量/密度
		POI分布	餐饮、景点、企业、购物等13类POI数量/密度
		监管属性	派出所及警务工作站数量/密度
时间特征	道路网络/行政区划	日历信息	月份、日期、星期、节假日、小时
外部特征	道路网络/行政区划	气象数据	风速、温度、气压、湿度、降水

模型类型	模型名称	关键超参数配置
机器学习	随机森林	n_estimators=[100,200]；max_depth=[None,10,20]；class_weight=[“balanced”]
机器学习	XGBoost	n_estimators=[100,200]；learning_rate=[0.01,0.1]；max_depth=[3,5,7]
深度学习	AFM	attention_factor=8；l2_reg_linear=1×10^-5；l2_reg_embedding=1×10^-5
	DCN	dnn_hidden_units=（256,128,64）；dnn_dropout=0.5；l2_reg_embedding=1×10^-4
	NFM	dnn_hidden_units=（256,128,64）；dnn_dropout=0.5；l2_reg_embedding=1×10^-4
	DeepFM	dnn_hidden_units=（256,128,64）；dnn_dropout=0.5；l2_reg_embedding=1×10^-4
图神经网络	DSTGCN	learning_rate=0.001；weight_decay=0.001
	DSTGAT	learning_rate=0.001；weight_decay=0.001
	MS-STGFN	learning_rate=0.001weight_decay=0.001

模型	所有警情		盗窃类警情		纠纷类警情		伤害类警情
模型	F1值	RMSE	F1值	RMSE	F1值	RMSE	F1值	RMSE
RF	0.6572	0.5022	0.6157	0.5233	0.5907	0.5341	0.5998	0.5184
XGBoost	0.6598	0.5003	0.5689	0.5731	0.5398	0.5911	0.5401	0.5927
AFM	0.6788	0.4723	0.6912	0.4681	0.6573	0.4729	0.6744	0.4796
DCN	0.6932	0.4762	0.7025	0.4684	0.6628	0.4738	0.6879	0.4643
NFM	0.6763	0.4603	0.6861	0.4634	0.6684	0.4802	0.6752	0.4768
DeepFM	0.6892	0.4521	0.6983	0.4532	0.6521	0.4569	0.6735	0.4534
DSTGCN	0.7597	0.4351	0.7551	0.4443	0.7342	0.4355	0.7518	0.4249
DSTGAT	0.7631	0.4211	0.7593	0.4355	0.7556	0.4429	0.7644	0.4286
MS-STGFN	0.7729	0.4008	0.7701	0.4156	0.7684	0.4193	0.7754	0.4013

模型	所有警情			盗窃类警情
模型	准确率	F1值	均方根误差	准确率	F1值	均方根误差
消融1	0.7266	0.7303	0.4472	0.7218	0.7243	0.4524
消融2	0.7583	0.7403	0.4305	0.7537	0.7358	0.4372
消融3	0.7502	0.7439	0.4311	0.7459	0.7387	0.4379
消融4	0.7763	0.7635	0.4131	0.7704	0.7586	0.4198
消融5	0.7476	0.7392	0.4401	0.7428	0.7345	0.4453
消融6	0.7689	0.7497	0.4216	0.7642	0.7449	0.4283
消融7	0.7737	0.7658	0.4284	0.7691	0.7608	0.4352
消融8	0.7956	0.7729	0.4008	0.7931	0.7701	0.4156
模型	纠纷类警情			伤害类警情
模型	准确率	F1值	均方根误差	准确率	F1值	均方根误差
消融1	0.7136	0.7164	0.4638	0.7307	0.7328	0.4426
消融2	0.7449	0.7268	0.4483	0.7616	0.7437	0.4274
消融3	0.7368	0.7305	0.4486	0.7538	0.7476	0.4283
消融4	0.7627	0.7496	0.4304	0.7796	0.7658	0.4103
消融5	0.7342	0.7258	0.4567	0.7514	0.7428	0.4369
消融6	0.7553	0.7362	0.4385	0.7729	0.7535	0.4182
消融7	0.7602	0.7519	0.4454	0.7775	0.7696	0.4253
消融8	0.7862	0.7684	0.4193	0.7993	0.7754	0.4013