Elsevier

Neurocomputing

Volume 486, 14 May 2022, Pages 286-297
Neurocomputing

Towards hour-level crime prediction: A neural attentive framework with spatial–temporal-categorical fusion

https://doi.org/10.1016/j.neucom.2021.11.052Get rights and content

Abstract

As one of the most complex social problems around the world, crime may bring the risk of dying or losing property to the public if not handled properly. Crime prediction which aims at predicting crime incidents before they happen is of great importance to fight against crime. Previous studies are concerned primarily with day-level crime prediction and have certain limitations on modeling complex spatial–temporal-categorical dependency contained in the criminal activities as well as utilizing external factors to facilitate the forecast. In this paper, we develop a novel Neural Attentive framework for Hour-level Crime prediction (NAHC) to cope with these challenges. Specifically, we first adopt the priori knowledge-based data enhancement strategy to alleviate the zero-inflated issue raised in hour-level settings. Then, multi-graph convolutional networks are applied to capture spatial dependency from different aspects. After that, we integrate gated recurrent units with a temporal attention mechanism to jointly address temporal dependency and capture time-sensitive external factors. A categorical attention mechanism is proposed for dealing with categorical dependency and finally a fully connected network is utilized to generate the final prediction results. Extensive experiments on two real-world crime datasets demonstrate the effectiveness of our framework over the state-of-the-art comparing methods.

Introduction

Crime has become one of the most troublesome problems in modern society, posing tremendous risks to human life and property. According to the annual report released by the Federal Bureau of Investigation, there were an estimated 1,247,321 violent crimes in the United States in 2017, including 17,284 incidents of murder, 135,755 rapes, 810,825 aggravated assaults, as well as 319,356 robberies.1 Furthermore, yearly financial losses due to burglary alone in the United States amount to 3.6 billion U.S. dollars, with an average cost of 2,361 dollars per incident.2 In view of such serious situation, governments around the world have put crime prevention high on their management agendas and call on everyone to take actions to fight against crime.

Evidently, crime prediction that aims at forecasting crime incidents ahead of time plays a vital role in crime prevention, which will not only assist urban authorities in taking effective measures to contain criminal activities, but also guide the public to keep away from zones with high potential risks. During the past decade, extensive efforts have been devoted to coping with this challenge and a series of approaches have been proposed. On the one hand, a number of researchers exploit heterogeneous data from multiple sources to facilitate the prediction task [1], [2], [3], [4], [5], [6]. For example, Gerber extracts discussion topics from Twitter messages and combines them with Kernel Density Estimation (KDE) to improve crime prediction performance [1]. On the other hand, some scholars focus on developing tailor-made algorithms for crime prediction task [7], [8], [9], [10], [11]. For instance, Zhao et al. propose a multi-task learning based framework called TCP to capture intra-region temporal correlations and inter-region spatial correlations simultaneously [7].

Although existing methods have remarkably improved the prediction performance, there still remain some issues which deserve careful consideration. Firstly, the majority of the methods are proposed for day-level prediction, i.e., forecasting how many crime incidents will happen in one day, which cannot satisfy the demands raised by many real-world applications. For instance, it would be nice if patrol routes could be arranged automatically for the police based on the prediction results. Considering that the risk level of a region varies with time, patrol routes should be scheduled differently for different time slots in a day, which calls for a method designed for providing hour-level prediction, i.e., predicting how many incidents will happen within the next few hours rather than one day. However, it is not a trivial matter to extend a day-level prediction method to hour-level, since the distribution of the crime incidents is extremely unbalanced where too many regions may have no incidents for most of the time [12], especially when the temporal resolution is set as hours. This phenomenon is called zero-inflated issue which will exert an adverse effect on the prediction results [13], [14], [15].

Secondly, crime incidents usually exhibit complex spatial, temporal and categorical dependency. Existing studies mostly assume that the number of crime incidents in one region is influenced by its nearby regions. However, two non-neighboring regions would have similar crime patterns as well if their functions are alike [9], [16]. Not only that, most of the studies pay their primary attention to short-term proximity, i.e., they assume that the number of crime incidents within a time slot is closely related to the number of crime incidents not far away from that slot [7], [8]. In addition to short-term proximity, criminal activities often have long-term periodicity which is overlooked by the majority of the models designed for crime prediction. Moreover, crime incidents falling into different categories may affect each other in an intricate manner. For example, a rise in the number of burglaries committed at one area may lead to the decrease of future criminal activities at the same location, which is a result of sending extra cops to strengthen patrols for areas with high risks [9]. Evidently, it is of great significance to utilize these three kinds of dependency comprehensively to facilitate the prediction process. Thirdly, lots of external factors may affect the criminal activities as well. For instance, crime is tightly linked to weather conditions [17], [18]. In addition, human leisure patterns on weekends could be quite different from their workday counterparts, leading the crime patterns on these two types of days vary considerably. Obviously, taking these factors into account is conducive to improving the performance.

To tackle the aforementioned challenges, we propose a novel Neural Attentive framework for Hour-level Crime prediction, denoted as NAHC. It not only makes full use of the spatial, temporal and categorical dependency between crime incidents but also fuses multiple external factors together to yield better prediction results. Specifically, we first adopt the Priori Knowledge-based Data Enhancement (PKDE) strategy which transforms zero values into negative values appropriately to alleviate the zero-inflated issue. Then, three kinds of adjacency graphs are constructed to reflect distance, functional similarity, and crime pattern similarity between two regions and Multi-Graph Convolutional Networks (MGCNs) are employed to capture spatial dependency from different angles by aggregating diverse information from multiple graphs. After that, we combine Gated Recurrent Units (GRUs) with a temporal attention mechanism to model short-term and long-term temporal dependency as well as capture time-sensitive external factors simultaneously. Subsequently, a categorical attention mechanism is designed for handling categorical dependency and finally a fully connected network is applied to generate the prediction results.

The reminder of this paper is organized as follows. In Section 2, we briefly revisit the work mostly related to our task. Then, we formally define the crime prediction problem in Section 3 and describe the details of our framework in Section 4. In Section 5 we present the experimental results and finally conclude our work in Section 6.

Section snippets

Related work

In this section, we give a brief overview of the work closely related to our study from three aspects, namely crime pattern analysis, crime prediction and spatiotemporal event prediction.

Preliminaries

In this section, we first elaborate some basic concepts and mathematical notations and then formally define the problem studied in this work.

Suppose there are I non-overlapping regions in a city and J crime categories (e.g., burglary, assault, and vehicle theft) over T time slots. We use a three-order tensor XRI×T×J to denote the historical crime data, where xitj is the number of crime incidents belong to the jth category committed at the ith region during the tth time slot. The length of one

Methodology

In this section, we elaborate the proposed deep learning based framework designed for hour-level crime prediction. It mainly consists of five parts which are used to addressing zero-inflated issue, modeling spatial, temporal and categorical dependency and generating prediction results respectively. Fig. 1 gives the overall architecture of our framework. Let us discuss in detail.

Experiments

In this section, we conduct extensive experiments on two real-world crime datasets to verify the effectiveness of our framework. We begin with an introduction to the experimental settings. Then, we present the performance of the proposed framework as well as each comparing algorithm. In particular, we aim to answer the following questions:

  • Q1: Compared with the state-of-the-art forecasting algorithms, can our framework achieve comparable prediction performance on both of the datasets in terms of

Conclusion

In this paper, we propose a novel neural attentive framework for hour-level crime prediction. Specifically, we first employ the priori knowledge-based data enhancement strategy to address zero-inflated issue. Then, we construct three kinds of graphs and utilize multi-graph convolution to capture spatial dependency from different angles. After that, we combine gated recurrent units with a temporal attention mechanism to model temporal dependency and capture external factors. Subsequently, we

CRediT authorship contribution statement

Weichao Liang: Conceptualization, Methodology, Software, Writing-original-draft, Writing-review-editing. Youquan Wang: Conceptualization, Methodology, Supervision. Haicheng Tao: Methodology, Investigation, Visualization. Jie Cao: Supervision, Project-administration, Funding-acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 92046026, Grant 72172057, and Grant 71701089, and in part by the Natural Science Foundation of the Higher Education Institutions of Jiangsu Province, China under Grant 21KJB520034.

Weichao Liang is currently pursuing the Ph.D. degree with the School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China. His main research interests include urban computing and data mining.

References (52)

  • L.G. Alves et al.

    Crime prediction through urban metrics and statistical learning

    Physica A

    (2018)
  • A. Bogomolov et al.

    Once upon a crime: towards crime prediction from demographics and mobile data

  • H. Wang et al.

    Crime rate inference with big data

  • H. Wang et al.

    Non-stationary model for crime rate inference using modern urban data

    IEEE Trans. Big Data

    (2019)
  • D. Yang et al.

    CrimeTelescope: crime hotspot prediction based on urban and social media data fusion

    World Wide Web

    (2018)
  • F. Yi et al.

    An integrated model for crime prediction using temporal and spatial factors

  • C. Huang et al.

    DeepCrime: attentive hierarchical recurrent networks for crime prediction

  • F. Yi et al.

    Neural network based continuous conditional random field for fine-grained crime prediction

  • C. Huang, C. Zhang, J. Zhao, X. Wu, D. Yin, N. Chawla, MiST: a multiview and multimodal spatial-temporal learning...
  • Z. Zhou, Y. Wang, X. Xie, L. Chen, H. Liu, RiskOracle: a minute-level citywide traffic accident forecasting framework,...
  • B. Wang, Y. Lin, S. Guo, H. Wan, GSNet: Learning spatial-temporal correlations from geographical and semantic aspects...
  • A. Belesiotis, G. Papadakis, D. Skoutas, Analyzing and predicting spatial crime distribution using crowdsourced and...
  • E.G. Cohn

    Weather and crime

    Br. J. Criminol.

    (1990)
  • J.L. Toole, N. Eagle, J.B. Plotkin, Spatiotemporal correlations in criminal offense records, ACM Trans. Intell. Syst....
  • C. Chauhan, S. Sehgal, A review: crime analysis using data mining techniques and algorithms, in: Proceedings of the 3rd...
  • R.A. Adeyemi et al.

    Demography and crime: a spatial analysis of geographical patterns and risk factors of crimes in nigeria

    Spatial Stat.

    (2021)
  • Cited by (11)

    View all citing articles on Scopus

    Weichao Liang is currently pursuing the Ph.D. degree with the School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China. His main research interests include urban computing and data mining.

    Youquan Wang received the Ph.D. degree in computer application technology from Nanjing University of Science and Technology, Nanjing, China. He is currently a Lecturer with the Jiangsu Provincial Key Laboratory of E-Business, Nanjing University of Finance and Economics, Nanjing, China. He has published more than 30 refereed journal and conference papers in these areas. He is the Member of the ACM and CCF. His research interests include deep learning and data mining.

    Haicheng Tao received the Ph.D. degree in computer science from Nanjing University of Science and Technology, Nanjing, China. He is currently a Lecturer with Nanjing University of Finance and Economics, Nanjing, China. His research interests include data mining, machine learning, artificial intelligence and optimization.

    Jie Cao received the Ph.D. degree in information science and engineering from Southeast University, Nanjing, China, in 2002. He is currently a Professor and the Dean of School of Information Engineering, Nanjing University of Finance and Economics, Nanjing, China. His main research interests include data mining, deep learning, and business intelligence. He has been selected in the Program for New Century Excellent Talents in University and was the recipient of Young and Mid-Aged Expert with Outstanding Contribution in Jiangsu province.

    View full text