crime prediction research paper

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Journal Proposal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

A systematic review of multi-scale spatio-temporal crime prediction methods, 1. introduction.

• Studies related to crime prediction are systematically reviewed from various temporal and spatial perspectives.
• Common temporal and spatial crime prediction methods and evaluation metrics are summarized.
• The limitations of the current study are reviewed, and reasonable suggestions for future directions of exploration are provided.

2. Materials and Methods

2.1. publications sources, 2.1.1. publications search, 2.1.2. publications screening, 2.2. research overview, 2.2.1. research content, 2.2.2. prediction methods, 2.2.3. types of crime predicted, 2.2.4. evaluation metrics, 3. crime prediction methods and evaluation metrics, 3.1. crime prediction methods, 3.1.1. machine learning, 3.1.2. crime mapping, 3.1.3. other prediction methods, 3.2. evaluation metrics, 3.2.1. hit rate, 3.2.2. pai and pei, 3.2.3. accuracy and f1 score, 3.2.4. roc and auc, 3.2.5. mse and rmse, 3.2.6. r 2 and adjusted r 2, 4. temporal crime prediction, 4.1. temporal crime prediction based on crime data only, 4.1.1. short-term prediction, 4.1.2. medium-term and long-term prediction, 4.2. temporal crime prediction based on crime and external data, 4.2.1. short-term prediction, 4.2.2. medium-term prediction, 4.2.3. long-term prediction, 4.3. limitations of temporal prediction research.

• Data sparsity. Despite the advancements in improving the accuracy of the models, most prediction models are driven by data and still have difficulties in dealing with data sparsity. Some study areas have limited crime data, making it challenging to support crime prediction. Furthermore, as the granularity of time and space becomes finer, the data become sparser and the amount of irrelevant information gradually increases, leading to difficulties in modeling crime. It also exposes issues regarding the difficulty involved in using data-driven models to accurately identify and extract crime-related features. Adding external features may result in reduced correlation between data and crime or even the phenomenon of the “Curse of Dimensionality”, where the model cannot converge quickly in a short time.
• Insufficient practicality, interpretability, and transparency of the model. ML-based prediction models often lack interpretability due to their “black-box” nature. The improved performance of the model comes at the cost of interpretability. As the complexity of the model increases, its performance becomes stronger, but its interpretability becomes worse. It is not enough to evaluate a model based on accuracy alone; understanding the mechanics behind how the model works is crucial. It is important to know how prediction results are given and which features are crucial for the model, among other considerations. Otherwise, full trust in the prediction results cannot be established. Thus, there is a strong need to introduce model interpretability methods to improve the understanding of how the models function. Additionally, since the crime situation varies between regions, models trained in one region may not necessarily transfer well to other regions.
• Single evaluation system. The evaluation metrics and data used in the above studies vary, making it impossible to judge the merits of the models accurately. Some studies rely solely on historical crime data, while others use a combination of crime and external data, such as demographic, socio-economic, and environmental factors. Moreover, the evaluation metrics are often too narrow, making it challenging to compare the performance of the models accurately. Thus, there is a need to establish a comprehensive evaluation system that considers various data types and evaluation metrics to truly judge the merit of the models.
• Limited studies on short-term crime prediction. Most of the studies discussed above focus primarily on medium- and long-term crime prediction (monthly, quarterly, and annual), which has a positive impact on macro-level policy making. However, few studies concentrate on short-term prediction at the hourly, daily, and weekly levels. Short-term prediction better serves the needs of most police departments since crimes such as burglary and robbery are typically short-lived, and they require rapid action to prevent and combat the crimes effectively. The lack of research in short-term prediction models makes it difficult to prevent crime from happening, such as by deploying officers and planning patrol routes aimed at targeted areas. When an offense occurs, the perpetrators cannot be caught in time, resulting in a significant blow to law enforcement.

5. Spatial Crime Prediction

5.1. micro- and meso-level prediction, 5.2. macro-level prediction, 5.3. limitations of spatial crime prediction research.

• Research on spatial crime prediction has made significant strides in identifying potential risk factors and crime hotspots, while also validating relevant criminology theories. However, despite the progress made thus far, this field still faces numerous challenges. For instance, the accuracy of crime prediction models heavily relies on the quality of crime data and the availability of relevant urban features. Moreover, the effective integration of various data sources remains a significant challenge in the development of reliable crime prediction models. There are fewer studies at the micro-level. While most studies in the spatial crime prediction field focus on macro-level predictions due to the availability of city-related data, they often overlook the importance of micro-level predictions. Conducting micro-level research would prove invaluable as it could assist the police in achieving scientific resource allocation and dispatch for specific areas and roads. Moreover, such research could enable enterprises to choose suitable business locations and to help citizens select safe travel routes and times, thus mitigating crime opportunities and promoting crime deterrence. The incorporation of micro-level predictions could provide more nuanced and context-specific insights and recommendations to a diverse group of stakeholders, thereby improving the effectiveness and efficiency of crime prevention strategies.
• Lack of research on decision-making applications. Some studies lack practical support for assisted decision making, which impedes their practical applications. Going forward, there should be a stronger emphasis on the implementation of research results in assisting the development of scientific police decisions. For instance, integrating patrol route planning research and other decision-making tools would be valuable in optimizing crime prevention efforts. By bridging the gap between research and practice, stakeholders can more effectively employ spatial crime prediction models for actionable insights and evidence-based decision making.
• Insufficient research on crime mechanisms. Although most spatial crime prediction studies successfully validate criminology-related theories and achieve the task of crime prediction to some extent, the directionality of some studies neglects the theoretical level. Consequently, these studies tend to only verify existing criminological theories, without sufficiently enriching or expanding the research on crime mechanisms. Additionally, the current research fails to consider the impact of criminal behavior patterns on crime prediction results. For instance, the presence of police on an offender’s travel route could deter the offender from committing the crime, which would subsequently affect the crime prediction accuracy. Therefore, future studies on spatial crime prediction should consider these contextual factors and aim to expand and advance crime mechanism research to improve the accuracy and applicability of crime prediction models.
• Unreasonable grid cell size. Most spatial crime prediction studies employ grid cells with side lengths of 100 m, 150 m, and 200 m. The research indicates that larger grid sizes generally result in better prediction performance. However, the theoretical limit range of a police patrol is 150 m, which should be considered when considering the relationship between grid size and police patrol range in practical crime prediction and police work. Flexibly adjusting grid size according to actual conditions and police patrol frequencies is essential. For areas with frequent police patrol, smaller grid cells should be used for prediction and analysis to improve prediction accuracy, while for areas with insufficient police patrol, larger grid cells should be employed to maximize the use of limited police resources and ensure comprehensive coverage. Striking a balance between prediction performance and practical application is key in optimizing the implementation of spatial crime prediction models.

6. Spatio-Temporal Crime Prediction

6.1. short-term and micro-level prediction, 6.2. short-term and meso-level prediction, 6.3. short-term and macro-level prediction, 6.4. medium-term and macro-level prediction, 6.5. long-term and meso-level prediction, 6.6. long-term and macro-level prediction, 6.7. limitations of spatio-temporal crime prediction research.

• Spatio-temporal correlation. Crime is influenced by various factors, such as time, environment, weather, and networks, resulting in strong spatio-temporal correlations that make it difficult for traditional machine learning and time series analysis models to fully capture local or global spatio-temporal correlations. Blindly adding spatio-temporal crime data to some studies may lead to the overfitting of the model.
• Spatio-temporal heterogeneity. The spatial and temporal distribution of crime is not uniform. Crime data in different times and regions often show differences, making it difficult for the same model to capture crime patterns in different times and regions simultaneously.

7. Conclusions and Future Perspectives

• The continuous development of big data technology has enabled the use of advanced machine learning, hotspot mapping, and other methods for precise spatio-temporal crime prediction, resulting in significant progress and breakthroughs in this field. However, it is important to acknowledge that some prediction methods and techniques may not be able to fully address the complex and dynamic nature of contemporary crime. Thus, based on the literature review and analysis, this paper proposes reasonable and practical solutions to address the pressing issues and challenges in the research on spatial crime prediction. By addressing these challenges, we can further optimize and improve the efficiency and applicability of spatial crime prediction models in real-world settings. Data sparsity can be dealt with using transfer learning technologies. Data sparsity is a common challenge in the crime prediction field; it can limit the accuracy and generalizability of prediction models. The use of transfer learning techniques, a novel machine learning approach, offers a viable solution to this problem. Transfer learning allows the application of information and knowledge from existing domains to related domains, thereby enabling the training of deep learning models to capture connections between data and avoid overfitting, even with limited crime data. Additionally, when dealing with more data features and larger dimensions, feature selection, feature extraction, and cross-validation methods can be employed to optimize the performance and efficiency of spatial crime prediction models. By incorporating these techniques, researchers can more effectively address data sparsity issues and enhance the ability of spatial crime prediction models to capture and extrapolate meaningful patterns and trends.
• Introducing model interpretability methods to improve the interpretability of models. Model interpretability is essential in enhancing the understanding and trustworthiness of prediction models. The issue of “black-box” models that produce predictions that are difficult to comprehend can be partially addressed by utilizing model interpretability methods, such as the LIME and SHAP models. LIME is a widely applicable model that facilitates both global and local interpretation of prediction results. On the other hand, the SHAP model considers all features as “contributors” and assigns SHAP values to each feature that are positively correlated with the contribution made by the variable. This approach enables the ranking of variables according to the SHAP value, thereby improving the interpretability of the model while retaining a high predictive performance. Incorporating model interpretability methods enhances our understanding of crime patterns and levels and enables the development of more scientific, accurate, timely, and effective prevention and control measures.
• Establishing a set of data use and evaluation systems for multiple scales. A standard dataset usage and model evaluation system is crucial for crime prediction studies to promote accuracy, consistency, and interoperability. To this end, it is recommended to establish a set of data use and evaluation systems at each scale. Firstly, the standardization of the data use is essential to facilitate model comparison and ensure that models with the same prediction objectives and requirements employ the same type of dataset. Secondly, developing a comprehensive evaluation system is vital to facilitate the accurate performance measurement of such models. Using consistent evaluation metrics for models with the same prediction objectives is critical in gauging the efficacy of different models and developing a cross-comparable model evaluation framework. By incorporating these measures, we can establish a standard data usage and model evaluation system that promotes the accuracy, validity, and practical viability of crime prediction models at multiple scales.
• Integrating other technologies to promote research in decision making. To address the current issues of low correlation between spatio-temporal crime prediction models and lack of targeted prevention and control strategies, innovative technologies such as crime simulation and reinforcement learning can be incorporated to enhance decision-making applications. With further advancements in crime simulation and reinforcement learning technologies, we can enhance the decision-making applications in the spatial crime prediction field and develop effective prevention and control strategies to combat crime more efficiently. Integrating spatio-temporal elements into a crime simulation model can provide a comprehensive approach to spatial crime prediction. Through crime simulation, the potential time and place of crime occurrences can be predicted, and the process of crime can be visually presented. Such information can be input into a deep reinforcement learning framework to optimize crime prevention and control strategies through a continuous learning process. The deep reinforcement learning (DRL) crime prevention and control strategy optimization model continuously learns and evaluates strategies for a wide range of crime scenarios, while selecting the optimal strategy for resource allocation in specific spatio-temporal environments [ 99 , 100 ]. Police agencies can utilize the reinforcement learning strategy selection mechanism to deploy police resources, develop patrol plans, and implement arrest operations effectively. In addition, this approach can also enable prompt apprehension of perpetrators and can minimize losses in the event of a crime occurrence. By combining simulation modeling, deep reinforcement learning, and crime prevention strategies, we can enhance the implementation and effectiveness of spatial crime prediction models, contributing to more efficient and targeted crime prevention and control measures.

Author Contributions

Data availability statement, conflicts of interest.

• Malleson, N. Using agent-based models to simulate crime. In Agent-Based Models of Geographical Systems ; Springer: Dordrecht, The Netherlands, 2011; pp. 411–434. [ Google Scholar ] [ CrossRef ]
• Brantingham, P.L.; Brantingham, P.J. Mobility, Notoriety, and Crime: A Study in the Crime Patterns of Urban Nodal Points. J. Environ. Syst. 1981 , 11 , 89–99. [ Google Scholar ] [ CrossRef ]
• Weisburd, D. The law of crime concentration and the criminology of place. Criminology 2015 , 53 , 133–157. [ Google Scholar ] [ CrossRef ]
• Curman, A.S.N.; Andresen, M.A.; Brantingham, P.J. Crime and Place: A Longitudinal Examination of Street Segment Patterns in Vancouver, BC. J. Quant. Criminol. 2015 , 31 , 127–147. [ Google Scholar ] [ CrossRef ]
• Ratcliffe, J.H. Crime Mapping and the Training Needs of Law Enforcement. Eur. J. Crim. Policy Res. 2002 , 10 , 65–83. [ Google Scholar ] [ CrossRef ]
• Wang, Q.; Jin, G.; Zhao, X.; Feng, Y.; Huang, J. CSAN: A neural network benchmark model for crime forecasting in spatio-temporal scale. Knowl.-Based Syst. 2020 , 189 , 105120. [ Google Scholar ] [ CrossRef ]
• Weisburd, D.; Braga, A.A.; Groff, E.R.; Wooditch, A. Can hot spots policing reduce crime in urban areas? An agent-based simulation. Criminology 2017 , 55 , 137–173. [ Google Scholar ] [ CrossRef ]
• Zhu, H.; Wang, F. An agent-based model for simulating urban crime with improved daily routines. Comput. Environ. Urban Syst. 2021 , 89 , 101680. [ Google Scholar ] [ CrossRef ]
• Liberati, A.; Altman, D.G.; Tetzlaff, J.; Mulrow, C.; Gøtzsche, P.C.; Ioannidis, J.P.; Clarke, M.; Devereaux, P.J.; Kleijnen, J.; Moher, D. The PRISMA Statement for Reporting Systematic Reviews and Meta-Analyses of Studies That Evaluate Health Care Interventions: Explanation and Elaboration. Ann. Intern. Med. 2009 , 151 , W65–W94. [ Google Scholar ] [ CrossRef ]
• Jordan, M.I.; Mitchell, T.M. Machine learning: Trends, perspectives, and prospects. Science 2015 , 349 , 255–260. [ Google Scholar ] [ CrossRef ]
• Liakos, K.G.; Busato, P.; Moshou, D.; Pearson, S.; Bochtis, D. Machine learning in agriculture: A review. Sensors 2018 , 18 , 2674. [ Google Scholar ] [ CrossRef ]
• Brunton, S.L.; Noack, B.R.; Koumoutsakos, P. Machine learning for fluid mechanics. Annu. Rev. Fluid Mech. 2020 , 52 , 477–508. [ Google Scholar ] [ CrossRef ]
• Bühlmann, P. Bagging, boosting and ensemble methods. In Handbook of Computational Statistics ; Springer: Berlin/Heidelberg, Germany, 2012; pp. 985–1022. [ Google Scholar ]
• Bentéjac, C.; Csörgő, A.; Martínez-Muñoz, G. A comparative analysis of gradient boosting algorithms. Artif. Intell. Rev. 2020 , 54 , 1937–1967. [ Google Scholar ] [ CrossRef ]
• Song, K.; Yan, F.; Ding, T.; Gao, L.; Lu, S. A steel property optimization model based on the XGBoost algorithm and improved PSO. Comput. Mater. Sci. 2019 , 174 , 109472. [ Google Scholar ] [ CrossRef ]
• Wang, B.; Zhang, D.; Zhang, D.; Brantingham, P.J.; Bertozzi, A.L. Deep learning for real time crime forecasting. arXiv 2017 , arXiv:1707.03340. [ Google Scholar ] [ CrossRef ]
• Kang, H.-W.; Kang, H.-B. Prediction of crime occurrence from multi-modal data using deep learning. PLoS ONE 2017 , 12 , e0176244. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Dong, Q.; Ye, R.; Li, G. Crime amount prediction based on 2D convolution and long short-term memory neural network. ETRI J. 2022 , 44 , 208–219. [ Google Scholar ] [ CrossRef ]
• Andresen, M.A.; Hodgkinson, T. Predicting Property Crime Risk: An Application of Risk Terrain Modeling in Vancouver, Canada. Eur. J. Crim. Policy Res. 2018 , 24 , 373–392. [ Google Scholar ] [ CrossRef ]
• Boppuru, P.R.; Ramesha, K. Geo-Spatial Crime Analysis Using Newsfeed Data in Indian Context. Int. J. Web-Based Learn. Teach. Technol. 2019 , 14 , 49–64. [ Google Scholar ] [ CrossRef ]
• Hart, T.; Zandbergen, P. Kernel density estimation and hotspot mapping: Examining the influence of interpolation method, grid cell size, and bandwidth on crime forecasting. Polic. Int. J. 2014 , 37 , 305–323. [ Google Scholar ] [ CrossRef ]
• Caplan, J.M.; Kennedy, L.W.; Miller, J. Risk Terrain Modeling: Brokering Criminological Theory and GIS Methods for Crime Forecasting. Justice Q. 2011 , 28 , 360–381. [ Google Scholar ] [ CrossRef ]
• Marchment, Z.; Gill, P. Systematic review and meta-analysis of risk terrain modelling (RTM) as a spatial forecasting method. Crime Sci. 2021 , 10 , 12. [ Google Scholar ] [ CrossRef ]
• Brantingham, P.; Brantingham, P. Criminality of place. Eur. J. Crim. Policy Res. 1995 , 3 , 5–26. [ Google Scholar ] [ CrossRef ]
• Kennedy, L.W.; Caplan, J.M.; Piza, E. Risk Clusters, Hotspots, and Spatial Intelligence: Risk Terrain Modeling as an Algorithm for Police Resource Allocation Strategies. J. Quant. Criminol. 2010 , 27 , 339–362. [ Google Scholar ] [ CrossRef ]
• Drawve, G.; Grubb, J.; Steinman, H.; Belongie, M. Enhancing Data-Driven Law Enforcement Efforts: Exploring how Risk Terrain Modeling and Conjunctive Analysis Fit in a Crime and Traffic Safety Framework. Am. J. Crim. Justice 2018 , 44 , 106–124. [ Google Scholar ] [ CrossRef ]
• Islam, K.; Raza, A. Forecasting crime using ARIMA model. arXiv 2020 , arXiv:2003.08006. [ Google Scholar ] [ CrossRef ]
• Nitta, G.R.; Rao, B.Y.; Sravani, T.; Ramakrishiah, N.; BalaAnand, M. LASSO-based feature selection and naïve Bayes classifier for crime prediction and its type. Serv. Oriented Comput. Appl. 2019 , 13 , 187–197. [ Google Scholar ] [ CrossRef ]
• Malleson, N.; Heppenstall, A.; See, L. Crime reduction through simulation: An agent-based model of burglary. Comput. Environ. Urban Syst. 2010 , 34 , 236–250. [ Google Scholar ] [ CrossRef ]
• Swaraj, A.; Verma, K.; Kaur, A.; Singh, G.; Kumar, A.; de Sales, L.M. Implementation of stacking based ARIMA model for prediction of COVID-19 cases in India. J. Biomed. Inform. 2021 , 121 , 103887. [ Google Scholar ] [ CrossRef ]
• Alabdulrazzaq, H.; Alenezi, M.N.; Rawajfih, Y.; Alghannam, B.A.; Al-Hassan, A.A.; Al-Anzi, F.S. On the accuracy of ARIMA based prediction of COVID-19 spread. Results Phys. 2021 , 27 , 104509. [ Google Scholar ] [ CrossRef ]
• Chainey, S.; Tompson, L.; Uhlig, S. The Utility of Hotspot Mapping for Predicting Spatial Patterns of Crime. Secur. J. 2008 , 21 , 4–28. [ Google Scholar ] [ CrossRef ]
• Kajita, M.; Kajita, S. Crime prediction by data-driven Green’s function method. Int. J. Forecast. 2019 , 36 , 480–488. [ Google Scholar ] [ CrossRef ]
• Farjami, Y.; Abdi, K. A genetic-fuzzy algorithm for spatio-temporal crime prediction. J. Ambient. Intell. Humaniz. Comput. 2021 , 1–13. [ Google Scholar ] [ CrossRef ]
• Langton, S.; Dixon, A.; Farrell, G. Six months in: Pandemic crime trends in England and Wales. Crime Sci. 2021 , 10 , 6. [ Google Scholar ] [ CrossRef ]
• Jha, S.; Yang, E.; Almagrabi, A.O.; Bashir, A.K.; Joshi, G.P. RETRACTED ARTICLE: Comparative analysis of time series model and machine testing systems for crime forecasting. Neural Comput. Appl. 2020 , 33 , 10621–10636. [ Google Scholar ] [ CrossRef ]
• Shoesmith, G.L. Space–time autoregressive models and forecasting national, regional and state crime rates. Int. J. Forecast. 2013 , 29 , 191–201. [ Google Scholar ] [ CrossRef ]
• Zhang, Y.; Siriaraya, P.; Kawai, Y.; Jatowt, A. Predicting time and location of future crimes with recommendation methods. Knowl.-Based Syst. 2020 , 210 , 106503. [ Google Scholar ] [ CrossRef ]
• Santos-Marquez, F. Spatial beta-convergence forecasting models: Evidence from municipal homicide rates in Colombia. J. Forecast. 2021 , 41 , 294–302. [ Google Scholar ] [ CrossRef ]
• Boivin, R. Routine activity, population(s) and crime: Spatial heterogeneity and conflicting Propositions about the neighborhood crime-population link. Appl. Geogr. 2018 , 95 , 79–87. [ Google Scholar ] [ CrossRef ]
• Altindag, D.T. Crime and unemployment: Evidence from Europe. Int. Rev. Law Econ. 2012 , 32 , 145–157. [ Google Scholar ] [ CrossRef ]
• Groot, W.; Brink, H.M.V.D. The effects of education on crime. Appl. Econ. 2010 , 42 , 279–289. [ Google Scholar ] [ CrossRef ]
• Jonathan, O.E.; Olusola, A.J.; Bernadin, T.C.A.; Inoussa, T.M. Impacts of Crime on Socio-Economic Development. Mediterr. J. Soc. Sci. 2021 , 12 , 71. [ Google Scholar ] [ CrossRef ]
• Ranson, M. Crime, weather, and climate change. J. Environ. Econ. Manag. 2014 , 67 , 274–302. [ Google Scholar ] [ CrossRef ]
• Inlow, A.R. A comprehensive review of quantitative research on crime, the built environment, land use, and physical geography. Sociol. Compass 2021 , 15 , e12889. [ Google Scholar ] [ CrossRef ]
• Vo, T.; Sharma, R.; Kumar, R.; Son, L.H.; Pham, B.T.; Bui, D.T.; Priyadarshini, I.; Sarkar, M.; Le, T. Crime rate detection using social media of different crime locations and Twitter part-of-speech tagger with Brown clustering. J. Intell. Fuzzy Syst. 2020 , 38 , 4287–4299. [ Google Scholar ] [ CrossRef ]
• Sypion-Dutkowska, N.; Leitner, M. Land Use Influencing the Spatial Distribution of Urban Crime: A Case Study of Szczecin, Poland. ISPRS Int. J. Geo-Inf. 2017 , 6 , 74. [ Google Scholar ] [ CrossRef ]
• Clancy, K.; Chudzik, J.; Snowden, A.J.; Guha, S. Reconciling data-driven crime analysis with human-centered algorithms. Cities 2022 , 124 , 103604. [ Google Scholar ] [ CrossRef ]
• Andresen, M.A. Unemployment, GDP, and Crime: The Importance of Multiple Measurements of the Economy. Can. J. Criminol. Crim. Justice 2015 , 57 , 35–58. [ Google Scholar ] [ CrossRef ]
• Hipp, J.R.; Bates, C.; Lichman, M.; Smyth, P. Using Social Media to Measure Temporal Ambient Population: Does it Help Explain Local Crime Rates? Justice Q. 2019 , 36 , 718–748. [ Google Scholar ] [ CrossRef ]
• Gerell, M. Does the Association Between Flows of People and Crime Differ Across Crime Types in Sweden? Eur. J. Crim. Policy Res. 2021 , 27 , 433–449. [ Google Scholar ] [ CrossRef ]
• Ding, N.; Zhai, Y. Crime prevention of bus pickpocketing in Beijing, China: Does air quality affect crime? Secur. J. 2019 , 34 , 262–277. [ Google Scholar ] [ CrossRef ]
• Venter, Z.S.; Shackleton, C.; Faull, A.; Lancaster, L.; Breetzke, G.; Edelstein, I. Is green space associated with reduced crime? A national-scale study from the Global South. Sci. Total. Environ. 2022 , 825 , 154005. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Hou, K.; Zhang, L.; Xu, X.; Yang, F.; Chen, B.; Hu, W.; Shu, R. High ambient temperatures are associated with urban crime risk in Chicago. Sci. Total. Environ. 2023 , 856 , 158846. [ Google Scholar ] [ CrossRef ]
• Ye, C.; Chen, Y.; Li, J. Investigating the Influences of Tree Coverage and Road Density on Property Crime. ISPRS Int. J. Geo-Inf. 2018 , 7 , 101. [ Google Scholar ] [ CrossRef ]
• Xu, Y.; Fu, C.; Kennedy, E.; Jiang, S.; Owusu-Agyemang, S. The impact of street lights on spatial-temporal patterns of crime in Detroit, Michigan. Cities 2018 , 79 , 45–52. [ Google Scholar ] [ CrossRef ]
• Ristea, A.; Al Boni, M.; Resch, B.; Gerber, M.S.; Leitner, M. Spatial crime distribution and prediction for sporting events using social media. Int. J. Geogr. Inf. Sci. 2020 , 34 , 1708–1739. [ Google Scholar ] [ CrossRef ]
• Stec, A.; Klabjan, D. Forecasting crime with deep learning. arXiv 2018 , arXiv:1806.01486. [ Google Scholar ] [ CrossRef ]
• Hu, J. A Hybrid GCN and LSTM Structure Based on Attention Mechanism for Crime Prediction. Converter 2021 , 328–338. [ Google Scholar ] [ CrossRef ]
• Han, X.; Hu, X.; Wu, H.; Shen, B.; Wu, J. Risk Prediction of Theft Crimes in Urban Communities: An Integrated Model of LSTM and ST-GCN. IEEE Access 2020 , 8 , 217222–217230. [ Google Scholar ] [ CrossRef ]
• Liang, X.; Liu, X.; Lan, M.; Song, G.; Xiao, L.; Chen, J. Interpretable machine learning models for crime prediction. Comput. Environ. Urban Syst. 2022 , 94 , 101789. [ Google Scholar ] [ CrossRef ]
• Liang, W.; Wang, Y.; Tao, H.; Cao, J. Towards hour-level crime prediction: A neural attentive framework with spatial–temporal-categorical fusion. Neurocomputing 2022 , 486 , 286–297. [ Google Scholar ] [ CrossRef ]
• Aghababaei, S.; Makrehchi, M. Mining Twitter data for crime trend prediction. Intell. Data Anal. 2018 , 22 , 117–141. [ Google Scholar ] [ CrossRef ]
• Huang, C.; Zhang, J.; Zheng, Y.; Chawla, N.V. DeepCrime: Attentive hierarchical recurrent networks for crime prediction. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, Torino, Italy, 22–26 October 2018; pp. 1423–1432. [ Google Scholar ]
• Rayhan, Y.; Hashem, T. AIST: An Interpretable Attention-Based Deep Learning Model for Crime Prediction. ACM Trans. Spat. Algorithms Syst. 2023 , 9 , 1–31. [ Google Scholar ] [ CrossRef ]
• Mahfoud, M.; Bernasco, W.; Bhulai, S.; van der Mei, R. Forecasting Spatio-Temporal Variation in Residential Burglary with the Integrated Laplace Approximation Framework: Effects of Crime Generators, Street Networks, and Prior Crimes. J. Quant. Criminol. 2020 , 37 , 835–862. [ Google Scholar ] [ CrossRef ]
• Wang, J.; Hu, J.; Shen, S.; Zhuang, J.; Ni, S. Crime risk analysis through big data algorithm with urban metrics. Phys. A Stat. Mech. Its Appl. 2020 , 545 , 123627. [ Google Scholar ] [ CrossRef ]
• Bappee, F.K.; Soares, A.; Petry, L.M.; Matwin, S. Examining the impact of cross-domain learning on crime prediction. J. Big Data 2021 , 8 , 96. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Kadar, C.; Pletikosa, I. Mining large-scale human mobility data for long-term crime prediction. EPJ Data Sci. 2018 , 7 , 26. [ Google Scholar ] [ CrossRef ]
• Alves, L.G.; Ribeiro, H.V.; Rodrigues, F.A. Crime prediction through urban metrics and statistical learning. Phys. A Stat. Mech. Its Appl. 2018 , 505 , 435–443. [ Google Scholar ] [ CrossRef ]
• Rummens, A.; Hardyns, W. The effect of spatio-temporal resolution on predictive policing model performance. Int. J. Forecast. 2021 , 37 , 125–133. [ Google Scholar ] [ CrossRef ]
• Zhang, Y.; Cheng, T. Graph deep learning model for network-based predictive hotspot mapping of sparse spatio-temporal events. Comput. Environ. Urban Syst. 2020 , 79 , 101403. [ Google Scholar ] [ CrossRef ]
• Rashidi, P.; Wang, T.; Skidmore, A.; Vrieling, A.; Darvishzadeh, R.; Toxopeus, B.; Ngene, S.; Omondi, P. Spatial and spatiotemporal clustering methods for detecting elephant poaching hotspots. Ecol. Model. 2015 , 297 , 180–186. [ Google Scholar ] [ CrossRef ]
• Drawve, G.; Thomas, S.A.; Walker, J.T. Bringing the physical environment back into neighborhood research: The utility of RTM for developing an aggregate neighborhood risk of crime measure. J. Crim. Justice 2016 , 44 , 21–29. [ Google Scholar ] [ CrossRef ]
• Drawve, G. A Metric Comparison of Predictive Hot Spot Techniques and RTM. Justice Q. 2014 , 33 , 369–397. [ Google Scholar ] [ CrossRef ]
• Wang, Z.; Liu, L.; Zhou, H.; Lan, M. Crime Geographical Displacement: Testing Its Potential Contribution to Crime Prediction. ISPRS Int. J. Geo-Inf. 2019 , 8 , 383. [ Google Scholar ] [ CrossRef ]
• Garnier, S.; Caplan, J.M.; Kennedy, L.W. Predicting Dynamical Crime Distribution From Environmental and Social Influences. Front. Appl. Math. Stat. 2018 , 4 , 13. [ Google Scholar ] [ CrossRef ]
• Rosés, R.; Kadar, C.; Malleson, N. A data-driven agent-based simulation to predict crime patterns in an urban environment. Comput. Environ. Urban Syst. 2021 , 89 , 101660. [ Google Scholar ] [ CrossRef ]
• Stalidis, P.; Semertzidis, T.; Daras, P. Examining Deep Learning Architectures for Crime Classification and Prediction. Forecasting 2021 , 3 , 741–762. [ Google Scholar ] [ CrossRef ]
• Saraiva, M.; Matijošaitienė, I.; Mishra, S.; Amante, A. Crime Prediction and Monitoring in Porto, Portugal, Using Machine Learning, Spatial and Text Analytics. ISPRS Int. J. Geo-Inf. 2022 , 11 , 400. [ Google Scholar ] [ CrossRef ]
• Belesiotis, A.; Papadakis, G.; Skoutas, D. Analyzing and Predicting Spatial Crime Distribution Using Crowdsourced and Open Data. ACM Trans. Spat. Algorithms Syst. 2018 , 3 , 1–31. [ Google Scholar ] [ CrossRef ]
• Dong, Q.; Li, Y.; Zheng, Z.; Wang, X.; Li, G. ST3DNetCrime: Improved ST-3DNet Model for Crime Prediction at Fine Spatial Temporal Scales. ISPRS Int. J. Geo-Inf. 2022 , 11 , 529. [ Google Scholar ] [ CrossRef ]
• Liu, L.; Ji, J.; Song, G.; Song, G.; Liao, W.; Yu, H.; Liu, W. Hotspot prediction of public property crime based on spatial differentiation of crime and built environment. J. Geo-Inf. Sci. 2019 , 21 , 1655–1668. [ Google Scholar ]
• Fitterer, J.; Nelson, T.; Nathoo, F. Predictive crime mapping. Police Pract. Res. 2014 , 16 , 121–135. [ Google Scholar ] [ CrossRef ]
• Hou, M.; Hu, X.; Cai, J.; Han, X.; Yuan, S. An Integrated Graph Model for Spatial–Temporal Urban Crime Prediction Based on Attention Mechanism. ISPRS Int. J. Geo-Inf. 2022 , 11 , 294. [ Google Scholar ] [ CrossRef ]
• Yu, H.; Liu, L.; Yang, B.; Lan, M. Crime Prediction with Historical Crime and Movement Data of Potential Offenders Using a Spatio-Temporal Cokriging Method. ISPRS Int. J. Geo-Inf. 2020 , 9 , 732. [ Google Scholar ] [ CrossRef ]
• Zhou, B.; Chen, L.; Zhou, F.; Li, S.; Zhao, S.; Das, S.K.; Pan, G. Escort: Fine-Grained Urban Crime Risk Inference Leveraging Heterogeneous Open Data. IEEE Syst. J. 2020 , 15 , 4656–4667. [ Google Scholar ] [ CrossRef ]
• Boppuru, P.R.; Ramesha, K. Spatio-Temporal Crime Analysis Using KDE and ARIMA Models in the Indian Context. Int. J. Digit. Crime Forensics 2020 , 12 , 1–19. [ Google Scholar ] [ CrossRef ]
• Zhao, X.; Tang, J. Modeling Temporal-Spatial Correlations for Crime Prediction. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, Singapore, 6–10 November 2017; pp. 497–506. [ Google Scholar ] [ CrossRef ]
• Kocher, M.; Leitner, M. Forecasting of crime events applying risk terrain modeling. GI_Forum–J. Geogr. Inf. 2015 , 1 , 30–40. [ Google Scholar ] [ CrossRef ]
• Rummens, A.; Hardyns, W.; Pauwels, L. The use of predictive analysis in spatiotemporal crime forecasting: Building and testing a model in an urban context. Appl. Geogr. 2017 , 86 , 255–261. [ Google Scholar ] [ CrossRef ]
• Hu, Y.; Wang, F.; Guin, C.; Zhu, H. A spatio-temporal kernel density estimation framework for predictive crime hotspot mapping and evaluation. Appl. Geogr. 2018 , 99 , 89–97. [ Google Scholar ] [ CrossRef ]
• Lin, Y.-L.; Yen, M.-F.; Yu, L.-C. Grid-Based Crime Prediction Using Geographical Features. ISPRS Int. J. Geo-Inf. 2018 , 7 , 298. [ Google Scholar ] [ CrossRef ]
• Solomon, A.; Kertis, M.; Shapira, B.; Rokach, L. A deep learning framework for predicting burglaries based on multiple contextual factors. Expert Syst. Appl. 2022 , 199 , 117042. [ Google Scholar ] [ CrossRef ]
• Law, J.; Quick, M.; Chan, P. Bayesian Spatio-Temporal Modeling for Analysing Local Patterns of Crime Over Time at the Small-Area Level. J. Quant. Criminol. 2014 , 30 , 57–78. [ Google Scholar ] [ CrossRef ]
• Kadar, C.; Maculan, R.; Feuerriegel, S. Public decision support for low population density areas: An imbalance-aware hyper-ensemble for spatio-temporal crime prediction. Decis. Support Syst. 2019 , 119 , 107–117. [ Google Scholar ] [ CrossRef ]
• Hu, T.; Zhu, X.; Duan, L.; Guo, W. Urban crime prediction based on spatio-temporal Bayesian model. PLoS ONE 2018 , 13 , e0206215. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Lamari, Y.; Freskura, B.; Abdessamad, A.; Eichberg, S.; De Bonviller, S. Predicting Spatial Crime Occurrences through an Efficient Ensemble-Learning Model. ISPRS Int. J. Geo-Inf. 2020 , 9 , 645. [ Google Scholar ] [ CrossRef ]
• Vimala Devi, J.; Kavitha, K.S. Adaptive deep Q learning network with reinforcement learning for crime prediction. Evolutionary Intelligence 2023 , 16 , 685–696. [ Google Scholar ] [ CrossRef ]
• Lim, M.; Abdullah, A.; Jhanjhi, N.Z.; Khan, M.K. Situation-aware deep reinforcement learning link prediction model for evolving criminal networks. IEEE Access 2019 , 8 , 16550–16559. [ Google Scholar ] [ CrossRef ]

Click here to enlarge figure

Share and Cite

Du, Y.; Ding, N. A Systematic Review of Multi-Scale Spatio-Temporal Crime Prediction Methods. ISPRS Int. J. Geo-Inf. 2023 , 12 , 209. https://doi.org/10.3390/ijgi12060209

Du Y, Ding N. A Systematic Review of Multi-Scale Spatio-Temporal Crime Prediction Methods. ISPRS International Journal of Geo-Information . 2023; 12(6):209. https://doi.org/10.3390/ijgi12060209

Du, Yingjie, and Ning Ding. 2023. "A Systematic Review of Multi-Scale Spatio-Temporal Crime Prediction Methods" ISPRS International Journal of Geo-Information 12, no. 6: 209. https://doi.org/10.3390/ijgi12060209

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

• Open access
• Published: 29 April 2021

Crime forecasting: a machine learning and computer vision approach to crime prediction and prevention

• Neil Shah 1 ,
• Nandish Bhagat 1 &
• Manan Shah   ORCID: orcid.org/0000-0002-8665-5010 2  

Visual Computing for Industry, Biomedicine, and Art volume  4 , Article number:  9 ( 2021 ) Cite this article

75k Accesses

70 Citations

4 Altmetric

Metrics details

A crime is a deliberate act that can cause physical or psychological harm, as well as property damage or loss, and can lead to punishment by a state or other authority according to the severity of the crime. The number and forms of criminal activities are increasing at an alarming rate, forcing agencies to develop efficient methods to take preventive measures. In the current scenario of rapidly increasing crime, traditional crime-solving techniques are unable to deliver results, being slow paced and less efficient. Thus, if we can come up with ways to predict crime, in detail, before it occurs, or come up with a “machine” that can assist police officers, it would lift the burden of police and help in preventing crimes. To achieve this, we suggest including machine learning (ML) and computer vision algorithms and techniques. In this paper, we describe the results of certain cases where such approaches were used, and which motivated us to pursue further research in this field. The main reason for the change in crime detection and prevention lies in the before and after statistical observations of the authorities using such techniques. The sole purpose of this study is to determine how a combination of ML and computer vision can be used by law agencies or authorities to detect, prevent, and solve crimes at a much more accurate and faster rate. In summary, ML and computer vision techniques can bring about an evolution in law agencies.

Introduction

Computer vision is a branch of artificial intelligence that trains the computer to understand and comprehend the visual world, and by doing so, creates a sense of understanding of a machine’s surroundings [ 1 , 2 ]. It mainly analyzes data of the surroundings from a camera, and thus its applications are significant. It can be used for face recognition, number plate recognition, augmented and mixed realities, location determination, and identifying objects [ 3 ]. Research is currently being conducted on the formation of mathematical techniques to recover and make it possible for computers to comprehend 3D images. Obtaining the 3D visuals of an object helps us with object detection, pedestrian detection, face recognition, Eigenfaces active appearance and 3D shape models, personal photo collections, instance recognition, geometric alignment, large databases, location recognition, category recognition, bag of words, part-based models, recognition with segmentation, intelligent photo editing, context and scene understanding, and large image collection and learning, image searches, recognition databases, and test sets. These are only basic applications, and each category mentioned above can be further explored. In ref. [ 4 ], VLFeat is introduced, which is a library of computer vision algorithms that can be used to conduct fast prototyping in computer vision research, thus enabling a tool to obtain computer vision results much faster than anticipated. Considering face detection/human recognition [ 5 ], human posture can also be recognized. Thus, computer vision is extremely attractive for visualizing the world around us.

Machine learning (ML) is an application that provides a system with the ability to learn and improve automatically from past experiences without being explicitly programmed [ 6 , 7 , 8 ]. After viewing the data, an exact pattern or information cannot always be determined [ 9 , 10 , 11 ]. In such cases, ML is applied to interpret the exact pattern and information [ 12 , 13 ]. ML pushes forward the idea that, by providing a machine with access to the right data, the machine can learn and solve both complex mathematical problems and some specific problems [ 14 , 15 , 16 , 17 ]. In general, ML is categorized into two parts: (1) supervised ML and (2) unsupervised ML [ 18 , 19 ]. In supervised learning, the machine is trained on the basis of a predefined set of training examples, which facilitates its capability to obtain precise and accurate conclusions when new data are given [ 20 , 21 ]. In unsupervised learning, the machine is given a set of data, and it must find some common patterns and relationships between the data its own [ 22 , 23 ]. Neural networks, which are important tools used in supervised learning, have been studied since the 1980s [ 24 , 25 ]. In ref. [ 26 ], the author suggested that different aspects are needed to obtain an exit from nondeterministic polynomial (NP)-completeness, and architectural constraints are insufficient. However, in ref. [ 27 ], it was proved that NP-completeness problems can be extended to neural networks using sigmoid functions. Although such research has attempted to demonstrate the various aspects of new ML approaches, how accurate are the results [ 28 , 29 , 30 ]?

Although various crimes and their underlying nature seem to be unpredictable, how unforeseeable are they? In ref. [ 31 ], the authors pointed out that as society and the economy results in new types of crimes, the need for a prediction system has grown. In ref. [ 32 ], crime trends and prediction technology called Mahanolobis and a dynamic time wrapping technique are given, delivering the possibility of predicting crime and apprehending the actual culprit. As described in ref. [ 33 ], in 1998, the United States National Institute of Justice granted five grants for crime forecasting as an extension to crime mapping. Applications of crime forecasting are currently being used by law enforcement in the United States, the United Kingdom, the Netherlands, Germany, and Switzerland [ 34 ]. Nowadays, criminal intellect with the help of advances in technology is improving with each passing year. Consequently, it has become necessary for us to provide the police department and the government with the means of a new and powerful machine (a set of programs) that can help them in their process of solving crimes. The main aim of crime forecasting is to predict crimes before they occur, and thus, the importance of using crime forecasting methods is extremely clear. Furthermore, the prediction of crimes can sometimes be crucial because it may potentially save the life of a victim, prevent lifelong trauma, and avoid damage to private property. It may even be used to predict possible terrorist crimes and activities. Finally, if we implement predictive policing with a considerable level of accuracy, governments can apply other primary resources such as police manpower, detectives, and funds in other fields of crime solving, thereby curbing the problem of crime with double the power.

In this paper, we aim to make an impact by using both ML algorithms and computer vision methods to predict both the nature of a crime and possibly pinpoint a culprit. Beforehand, we questioned whether the nature of the crime was predictable. Although it might seem impossible from the outside, categorizing every aspect of a crime is quite possible. We have all heard that every criminal has a motive. That is, if we use motive as a judgment for the nature of a crime, we may be able to achieve a list of ways in which crimes can be categorized. Herein, we discuss a theory where we combine ML algorithms to act as a database for all recorded crimes in terms of category, along with providing visual knowledge of the surroundings through computer vision techniques, and using the knowledge of such data, we may predict a crime before it occurs.

Present technological used in crime detection and prediction

Crime forecasting refers to the basic process of predicting crimes before they occur. Tools are needed to predict a crime before it occurs. Currently, there are tools used by police to assist in specific tasks such as listening in on a suspect’s phone call or using a body cam to record some unusual illegal activity. Below we list some such tools to better understand where they might stand with additional technological assistance.

One good way of tracking phones is through the use of a stingray [ 35 ], which is a new frontier in police surveillance and can be used to pinpoint a cellphone location by mimicking cellphone towers and broadcasting the signals to trick cellphones within the vicinity to transmit their location and other information. An argument against the usage of stingrays in the United States is that it violates the fourth amendment. This technology is used in 23 states and in the district of Columbia. In ref. [ 36 ], the authors provide insight on how this is more than just a surveillance system, raising concerns about privacy violations. In addition, the Federal Communicatons Commission became involved and ultimately urged the manufacturer to meet two conditions in exchange for a grant: (1) “The marketing and sale of these devices shall be limited to federal, state, local public safety and law enforcement officials only” and (2) “State and local law enforcement agencies must advance coordinate with the FBI the acquisition and use of the equipment authorized under this authorization.” Although its use is worthwhile, its implementation remains extremely controversial.

A very popular method that has been in practice since the inception of surveillance is “the stakeout”. A stakeout is the most frequently practiced surveillance technique among police officers and is used to gather information on all types of suspects. In ref. [ 37 ], the authors discuss the importance of a stakeout by stating that police officers witness an extensive range of events about which they are required to write a report. Such criminal acts are observed during stakeouts or patrols; observations of weapons, drugs, and other evidence during house searches; and descriptions of their own behavior and that of the suspect during arrest. Stakeouts are extremely useful, and are considered 100% reliable, with the police themselves observing the notable proceedings. However, are they actually 100% accurate? All officers are humans, and all humans are subject to fatigue. The major objective of a stakeout is to observe wrongful activities. Is there a tool that can substitute its use? We will discuss this point herein.

Another way to conduct surveillance is by using drones, which help in various fields such as mapping cities, chasing suspects, investigating crime scenes and accidents, traffic management and flow, and search and rescue after a disaster. In ref. [ 38 ], legal issues regarding the use of drones and airspace distribution problems are described. Legal issues include the privacy concerns raised by the public, with the police gaining increasing power and authority. Airspace distribution raises concerns about how high a drone is allowed to go.

Other surveillance methods include face recognition, license plate recognition, and body cams. In ref. [ 39 ], the authors indicated that facial recognition can be used to obtain the profile of suspects and analyze it from different databases to obtain more information. Similarly, a license plate reader can be used to access data about a car possibly involved in a crime. They may even use body cams to see more than what the human eye can see, meaning that the reader observes everything a police officer sees and records it. Normally, when we see an object, we cannot recollect the complete image of it. In ref. [ 40 ], the impact of body cams was studied in terms of officer misconduct and domestic violence when the police are making an arrest. Body cams are thus being worn by patrol officers. In ref. [ 41 ], the authors also mentioned how protection against wrongful police practices is provided. However, the use of body cams does not stop here, as other primary reasons for having a body camera on at all times is to record the happenings in front of the wearer in hopes of record useful events during daily activities or during important operations.

Although each of these methods is effective, one point they share in common is that they all work individually, and while the police can use any of these approaches individually or concurrently, having a machine that is able to incorporate the positive aspects of all of these technologies would be highly beneficial.

ML techniques used in crime prediction

In ref. [ 42 ], a comparative study was carried out between violent crime patterns from the Communities and Crime Unnormalized Dataset versus actual crime statistical data using the open source data mining software Waikato Environment for Knowledge Analysis (WEKA). Three algorithms, namely, linear regression, additive regression, and decision stump, were implemented using the same finite set of features on communities and actual crime datasets. Test samples were randomly selected. The linear regression algorithm could handle randomness to a certain extent in the test samples and thus proved to be the best among all three selected algorithms. The scope of the project was to prove the efficiency and accuracy of ML algorithms in predicting violent crime patterns and other applications, such as determining criminal hotspots, creating criminal profiles, and learning criminal trends.

When considering WEKA [ 43 ], the integration of a new graphical interface called Knowledge Flow is possible, which can be used as a substitute for Internet Explorer. IT provides a more concentrated view of data mining in association with the process orientation, in which individual learning components (represented by java beans) are used graphically to show a certain flow of information. The authors then describe another graphical interface called an experimenter, which as the name suggests, is designed to compare the performance of multiple learning schemes on multiple data sets.

In ref. [ 34 ], the potential of applying a predictive analysis of crime forecasting in an urban context is studied. Three types of crime, namely, home burglary, street robbery, and battery, were aggregated into grids of 200 m × 250 m and retrospectively analyzed. Based on the crime data of the previous 3 years, an ensemble model was applied to synthesize the results of logistic regression and neural network models in order to obtain fortnightly and monthly predictions for the year 2014. The predictions were evaluated based on the direct hit rate, precision, and prediction index. The results of the fortnightly predictions indicate that by applying a predictive analysis methodology to the data, it is possible to obtain accurate predictions. They concluded that the results can be improved remarkably by comparing the fortnightly predictions with the monthly predictions with a separation between day and night.

In ref. [ 44 ], crime predictions were investigated based on ML. Crime data of the last 15 years in Vancouver (Canada) were analyzed for prediction. This machine-learning-based crime analysis involves the collection of data, data classification, identification of patterns, prediction, and visualization. K-nearest neighbor (KNN) and boosted decision tree algorithms were also implemented to analyze the crime dataset. In their study, a total of 560,000 crime datasets between 2003 and 2018 were analyzed, and crime prediction with an accuracy of between 39% and 44% was obtained by predicting the crime using ML algorithms. The accuracy was low as a prediction model, but the authors concluded that the accuracy can be increased or improved by tuning both the algorithms and crime data for specific applications.

In ref. [ 45 ], a ML approach is presented for the prediction of crime-related statistics in Philadelphia, United States. The problem was divided into three parts: determining whether the crime occurs, occurrence of crime and most likely crime. Algorithms such as logistic regression, KNN, ordinal regression, and tree methods were used to train the datasets to obtain detailed quantitative crime predictions with greater significance. They also presented a map for crime prediction with different crime categories in different areas of Philadelphia for a particular time period with different colors indicating each type of crime. Different types of crimes ranging from assaults to cyber fraud were included to match the general pattern of crime in Philadelphia for a particular interval of time. Their algorithm was able to predict whether a crime will occur with an astonishing 69% accuracy, as well as the number of crimes ranging from 1 to 32 with 47% accuracy.

In ref. [ 46 ], the authors analyzed a dataset consisting of several crimes and predicted the type of crime that may occur in the near future depending on various conditions. ML and data science techniques were used for crime prediction in a crime dataset from Chicago, United States. The crime dataset consists of information such as the crime location description, type of crime, date, time, and precise location coordinates. Different combinations of models, such as KNN classification, logistic regression, decision trees, random forest, a support vector machine (SVM), and Bayesian methods were tested, and the most accurate model was used for training. The KNN classification proved to be the best with an accuracy of approximately 0.787. They also used different graphs that helped in understanding the various characteristics of the crime dataset of Chicago. The main purpose of this paper is to provide an idea of how ML can be used by law enforcement agencies to predict, detect, and solve crime at a much better rate, which results in a reduction in crime.

In ref. [ 47 ], a graphical user interface-based prediction of crime rates using a ML approach is presented. The main focus of this study was to investigate machine-learning-based techniques with the best accuracy in predicting crime rates and explore its applicability with particular importance to the dataset. Supervised ML techniques were used to analyze the dataset to carry out data validation, data cleaning, and data visualization on the given dataset. The results of the different supervised ML algorithms were compared to predict the results. The proposed system consists of data collection, data preprocessing, construction of a predictive model, dataset training, dataset testing, and a comparison of algorithms, as shown in Fig.  1 . The aim of this study is to prove the effectiveness and accuracy of a ML algorithm for predicting violent crimes.

Dataflow diagram

In ref. [ 48 ], a feature-level data fusion method based on a deep neural network (DNN) is proposed to accurately predict crime occurrence by efficiently fusing multi-model data from several domains with environmental context information. The dataset consists of data from an online database of crime statistics from Chicago, demographic and meteorological data, and images. Crime prediction methods utilize several ML techniques, including a regression analysis, kernel density estimation (KDE), and SVM. Their approach mainly consisted of three phases: collection of data, analysis of the relationship between crime incidents and collected data using a statistical approach, and lastly, accurate prediction of crime occurrences. The DNN model consists of spatial features, temporal features, and environmental context. The SVM and KDE models had accuracies of 67.01% and 66.33%, respectively, whereas the proposed DNN model had an astonishing accuracy of 84.25%. The experimental results showed that the proposed DNN model was more accurate in predicting crime occurrences than the other prediction models.

In ref. [ 49 ], the authors mainly focused on the analysis and design of ML algorithms to reduce crime rates in India. ML techniques were applied to a large set of data to determine the pattern relations between them. The research was mainly based on providing a prediction of crime that might occur based on the occurrence of previous crime locations, as shown in Fig.  2 . Techniques such as Bayesian neural networks, the Levenberg Marquardt algorithm, and a scaled algorithm were used to analyze and interpret the data, among which the scaled algorithm gave the best result in comparison with the other two techniques. A statistical analysis based on the correlation, analysis of variance, and graphs proved that with the help of the scaled algorithm, the crime rate can be reduced by 78%, implying an accuracy of 0.78.

Functionality of proposed approach

In ref. [ 50 ], a system is proposed that predicts crime by analyzing a dataset containing records of previously committed crimes and their patterns. The proposed system works mainly on two ML algorithms: a decision tree and KNN. Techniques such as the random forest algorithm and Adaptive Boosting were used to increase the accuracy of the prediction model. To obtain better results for the model, the crimes were divided into frequent and rare classes. The frequent class consisted of the most frequent crimes, whereas the rare class consisted of the least frequent crimes. The proposed system was fed with criminal activity data for a 12-year period in San Francisco, United States. Using undersampling and oversampling methods along with the random forest algorithm, the accuracy was surprisingly increased to 99.16%.

In ref. [ 51 ], a detailed study on crime classification and prediction using ML and deep learning architectures is presented. Certain ML methodologies, such as random forest, naïve Bayes, and an SVM have been used in the literature to predict the number of crimes and hotspot prediction. Deep learning is a ML approach that can overcome the limitations of some machine-learning methodologies by extracting the features from the raw data. This paper presents three fundamental deep learning configurations for crime prediction: (1) spatial and temporal patterns, (2) temporal and spatial patterns, and (3) spatial and temporal patterns in parallel. Moreover, the proposed model was compared with 10 state-of-the-art algorithms on 5 different crime prediction datasets with more than 10 years of crime data.

In ref. [ 52 ], a big data and ML technique for behavior analysis and crime prediction is presented. This paper discusses the tracking of information using big data, different data collection approaches, and the last phase of crime prediction using ML techniques based on data collection and analysis. A predictive analysis was conducted through ML using RapidMiner by processing historical crime patterns. The research was mainly conducted in four phases: data collection, data preparation, data analysis, and data visualization. It was concluded that big data is a suitable framework for analyzing crime data because it can provide a high throughput and fault tolerance, analyze extremely large datasets, and generate reliable results, whereas the ML based naïve Bayes algorithm can achieve better predictions using the available datasets.

In ref. [ 53 ], various data mining and ML technologies used in criminal investigations are demonstrated. The contribution of this study is highlighting the methodologies used in crime data analytics. Various ML methods, such as a KNN, SVM, naïve Bayes, and clustering, were used for the classification, understanding, and analysis of datasets based on predefined conditions. By understanding and analyzing the data available in the crime record, the type of crime and the hotspot of future criminal activities can be determined. The proposed model was designed to perform various operations such as feature selection, clustering, analysis, prediction, and evaluation of the given datasets. This research proves the necessity of ML techniques for predicting and analyzing criminal activities.

In ref. [ 54 ], the authors incorporated the concept of a grid-based crime prediction model and established a range of spatial-temporal features based on 84 types of geographic locations for a city in Taiwan. The concept uses ML algorithms to learn the patterns and predict crime for the following month for each grid. Among the many ML methods applied, the best model was found to be a DNN. The main contribution of this study is the use of the most recent ML techniques, including the concept of feature learning. In addition, the testing of crime displacement also showed that the proposed model design outperformed the baseline.

In ref. [ 55 ], the authors considered the development of a crime prediction model using the decision tree (J48) algorithm. When applied in the context of law enforcement and intelligence analysis, J48 holds the promise of mollifying crime rates and is considered the most efficient ML algorithm for the prediction of crime data in the related literature. The J48 classifier was developed using the WEKA tool kit and later trained on a preprocessed crime dataset. The experimental results of the J48 algorithm predicted the unknown category of crime data with an accuracy of 94.25287%. With such high accuracy, it is fair to count on the system for future crime predictions.

Comparative study of different forecasting methods

First, in refs. [ 56 , 57 ], the authors predicted crime using the KNNs algorithm in the years 2014 and 2013, respectively. Sun et al. [ 56 ] proved that a higher crime prediction accuracy can be obtained by combining the grey correlation analysis based on new weighted KNN (GBWKNN) filling algorithm with the KNN classification algorithm. Using the proposed algorithm, we were able to obtain an accuracy of approximately 67%. By contrast, Shojaee et al. [ 57 ] divided crime data into two parts, namely, critical and non-critical, and applied a simple KNN algorithm. They achieved an astonishing accuracy of approximately 87%.

Second, in refs. [ 58 , 59 ], crime is predicted using a decision tree algorithm for the years 2015 and 2013, respectively. In their study, Obuandike et al. [ 58 ] used the ZeroR algorithm along with a decision tree but failed to achieve an accuracy of above 60%. In addition, Iqbal et al. [ 59 ] achieved a stunning accuracy of 84% using a decision tree algorithm. In both cases, however, a small change in the data could lead to a large change in the structure.

Third, in refs. [ 60 , 61 ], a novel crime detection technique called naïve Bayes was implemented for crime prediction and analysis. Jangra and Kalsi [ 60 ] achieved an astounding crime prediction accuracy of 87%, but could not apply their approach to datasets with a large number of features. By contrast, Wibowo and Oesman [ 61 ] achieved an accuracy of only 66% in predicting crimes and failed to consider the computational speed, robustness, and scalability.

Below, we summarize the above comparison and add other models to further illustrate this comparative study and the accuracy of some frequently used models (Table  1 ).

Computer vision models combined with machine and deep learning techniques

In ref. [ 66 ], the study focused on three main questions. First, the authors question whether computer vision algorithms actually work. They stated that the accuracy of the prediction is 90% over fewer complex datasets, but the accuracy drops to 60% over complex datasets. Another concern we need to focus on is reducing the storage and computational costs. Second, they question whether it is effective for policing. They determined that a distinct activity detection is difficult, and pinpointed a key component, the Public Safety Visual Analytics Workstation, which includes many capabilities ranging from detection and localization of objects in camera feeds to labeling actions and events associated with training data, and allowing query-based searches for specific events in videos. By doing so, they aim to view every event as a computer-vision trained, recognized, and labeled event. The third and final question they ask is whether computer vision impacts the criminal justice system. The answer to this from their end is quite optimistic to say the least, although they wish to implement computer vision alone, which we suspect is unsatisfactory.

In ref. [ 67 ], a framework for multi-camera video surveillance is presented. The framework is designed so efficiently that it performs all three major activities of a typical police “stake-out”, i.e., detection, representation, and recognition. The detection part mixes video streams from multiple cameras to efficiently and reliably extract motion trajectories from videos. The representation helps in concluding the raw trajectory data to construct hierarchical, invariant, and content-rich descriptions of motion events. Finally, the recognition part deals with event classification (such as robbery and possibly murder and molestation, among others) and identification of the data descriptors. For an effective recognition, they developed a sequence-alignment kernel function to perform sequence data learning to identify suspicious/possible crime events.

In ref. [ 68 ], a method is suggested for identifying people for surveillance with the help of a new feature called soft biometry, which includes a person’s height, built, skin tone, shirt and trouser color, motion pattern, and trajectory history to identify and track passengers, which further helps in predicting crime activities. They have gone further and discussed some absurd human error incidents that have resulted in the perpetrators getting away. They also conducted experiments, the results of which were quite astounding. In one case, the camera catches people giving piggyback rides in more than one frame of a single shot video. The second scenario shows the camera’s ability to distinguish between airport guards and passengers.

In ref. [ 69 ], the authors discussed automated visual surveillance in a realistic scenario and used Knight, which is a multiple camera surveillance and monitoring system. Their major targets were to analyze the detection, tracking, and classification performances. The detection, tracking, and classification accuracies were 97.4%, 96.7%, and 88%, respectively. The authors also pointed to the major difficulties of illumination changes, camouflage, uninteresting moving objects, and shadows. This research again proves the reliability of computer vision models.

It is well known that an ideal scenario for a camera to achieve a perfect resolution is not possible. In ref. [ 70 ], security surveillance systems often produce poor-quality video, which could be a hurdle in gathering forensic evidence. They examined the ability of subjects to identify targeted individuals captured by a commercially available video security device. In the first experiment, subjects personally familiar with the targets performed extremely well at identifying them, whereas subjects unfamiliar with the targets performed quite poorly. Although these results might not seem to be very conclusive and efficient, police officers with experience in forensic identification performed as poorly as other subjects unfamiliar with the targets. In the second experiment, they asked how familiar subjects could perform so well, and then used the same video device edited clips to obscure the head, body, or gait of the targets. Hiding the body or gait produced a small decrease in recognition performance. Hiding the target heads had a dramatic effect on the subject’s ability to recognize the targets. This indicates that even if the quality of the video is low, the head the target was seen and recognized.

In ref. [ 71 ], an automatic number plate recognition (ANPR) model is proposed. The authors described it as an “image processing innovation”. The ANPR system consists of the following steps: (1) vehicle image capture, (2) preprocessing, (3) number plate extraction, (4) character segmentation, and (5) character recognition. Before the main image processing, a pre-processing of the captured image is conducted, which includes converting the red, green and blue image into a gray image, clamor evacuation, and border enhancement for brightness. The plate is then separated by judging its size. In character segmentation, the letters and numbers are separated and viewed individually. In character recognition, optical character recognition is applied to a given database.

Although real-time crime forecasting is vital, it is extremely difficult to achieve in practice. No known physical models provide a reasonable approximation with dependable results for such a complex system. In ref. [ 72 ], the authors adapted a spatial temporal residual network to well-represented data to predict the distribution of crime in Los Angeles at an hourly scale in neighborhood-sized parcels. These experiments were compared with several existing approaches for prediction, demonstrating the superiority of the proposed model in terms of accuracy. They compared their deep learning approach to ARIMA, KNN, and the historical average. In addition, they presented a ternarization technique to address the concerns of resource consumption for deployment in the real world.

In ref. [ 73 ], the authors conducted a significant study on crime prediction and showed the importance of non-crime data. The major objective of this research was taking advantage of DNNs to achieve crime prediction in a fine-grain city partition. They made predictions using Chicago and Portland crime data, which were further augmented with additional datasets covering the weather, census data, and public transportation. In the paper they split each city into grid cells (beats for Chicago and square grid for Portland). The crime numbers are broken into 10 bins, and their model predicts the most likely bin for each spatial region at a daily level. They train these data using increasingly complex neural network structures, including variations that are suited to the spatial and temporal aspects of the crime prediction problem. Using their model, they were able to predict the correct bin for the overall number of crimes with an accuracy of 75.6% for Chicago and 65.3% for Portland. They showed that adding the value of additional non-crime data was an important factor. They found that days with higher amounts of precipitation and snow decreased the accuracy of the model slightly. Then, considering the impact of transportation, the bus routes and train routes were presented within their beats, and it was shown that the beat containing a train station is on average 1.2% higher than its neighboring beats. The accuracy of a beat that contained one or more train lines passing through it was 0.5% more accurate than its neighboring beats.

In ref. [ 74 ], the authors taught a system how to monitor traffic and identify vehicles at night. They used the bright spots of the headlights and tail lights to identify an object first as a vehicle, and the bright light is extracted with a segmentation process, and then processed by a spatial clustering and tracking procedure that locates and analyzes the spatial and temporal features of the vehicle light. They also conducted an experiment in which, for a span of 20 min, the detection scores for cars and bikes were 98.79% and 96.84%, respectively. In another part of the test, they conducted the same test under the same conditions for 50 min, and the detection scores for cars and bikes were 97.58% and 98.48%, respectively. It is good for machines to be built at such a beginning level. This technology can also be used to conduct surveillance at night.

In ref. [ 75 ], an important approach for human motion analysis is discussed. The author mentions that human motion analysis is difficult because appearances are extremely variable, and thus stresses that focusing on marker-less vision-based human motion analysis has the potential to provide a non-obtrusive solution for the evaluation of body poses. The author claims that this technology can have vast applications such as surveillance, human-computer interaction, and automatic annotation, and will thus benefit from a robust solution. In this paper, the characteristics of human motion analysis are discussed. We divide the analysis part into two aspects, modeling and an estimation phase. The modeling phase includes the construction of the likelihood function [including the camera model, image descriptors, human body model and matching function, and (physical) constraints], and the estimation phase is concerned with finding the most likely pose given the likelihood (function result) of the surface. We discuss the model-free approaches separately.

In ref. [ 76 ], the authors provided insight into how we can achieve crime mapping using satellites. The need for manual data collection for mapping is costly and time consuming. By contrast, satellite imagery is becoming a great alternative. In this paper, they attempted to investigate the use of deep learning to predict crime rates directly from raw satellite imagery. They trained a deep convolutional neural network (CNN) on satellite images obtained from over 1 million crime-incident reports (15 years of data) collected by the Chicago Police Department. The best performing model predicted crime rates from raw satellite imagery with an astounding accuracy of 79%. To make their research more thorough, they conducted a test for reusability, and used the tested and learned Chicago models for prediction in the cities of Denver and San Francisco. Compared to maps made from years of data collected by the corresponding police departments, their maps have an accuracy of 72% and 70%, respectively. They concluded the following: (1) Visual features contained in satellite imagery can be successfully used as a proxy indicator of crime rates; (2) ConvNets are capable of learning models for crime rate prediction from satellite imagery; (3) Once deep models are used and learned, they can be reused across different cities.

In ref. [ 77 ], the authors suggested an extremely intriguing research approach in which they claim to prove that looking beyond what is visible is to infer meaning to what is viewed from an image. They even conducted an interesting study on determining where a McDonalds could be located simply from photographs, and provided the possibility of predicting crime. They compared the human accuracy on this task, which was 59.6%, and the accuracy of using gradient-based features, which was 72.5%, with a chance performance (a chance performance is what you would obtain if you performed at random) of only 50%. This indicates the presence of some visual cues that are not easily spotted by an average human, but are able to be spotted by a machine, thus enables us to judge whether an area is safe. The authors indicated that numerous factors are often associated with our intuition, which we use to avoid certain areas because they may seem “shady” or “unsafe”.

In ref. [ 78 ], the authors describe in two parts how close we are to achieving a fully automated surveillance system. The first part views the possibility of surveillance in a real-world scenario where the installation of systems and maintenance of systems are in question. The second part considers the implementation of computer vision models and algorithms for behavior modeling and event detection. They concluded that the complete scenario is under discussion, and therefore many people are conducting research and obtaining results. However, as we look closely, we can see that reliable results are possible only in certain aspects, while other areas are still in the development process, such as obtaining information on cars and their owners as well as accurately understanding the behavior of a possible suspect.

Many times during criminal activities, convicts use hand gestures to signal messages to each other. In ref. [ 79 ], research on hand gesture recognition was conducted using computer vision models. Their application architecture is of extremely high quality and is easy to understand. They begin by capturing images, and then try detecting a hand in the background. They apply either computer aided manufacturing or different procedure in which they first convert a picture into gray scale, after which they set the image return on investment, and then find and extract the biggest contour. They then determine the convex hull of the contour to try and find an orientation around the bounded rectangle, and finally interpret the gesture and convert it into a meaningful command.

Crime hotspots or areas with high crime intensity are places where the future possibility of a crime exists along with the possibility of spotting a criminal. In ref. [ 80 ], the authors conducted research on forecasting crime hotspots. They used Google Tensor Flow to implement their model and evaluated three options for the recurrent neural network (RNN) architecture: accuracy, precision, and recall. The focus is on achieving a larger value to prove that the approach has a better performance. The gated recurrent unit (GRU) and long short-term memory (LSTM) versions obtained similar performance levels with an accuracy of 81.5%, precision of 86%–87%, recall of 75%, and F1-score of 0.8. Both perform much better than the traditional RNN version. Based on the area under the ROC curve (AUC) performance observations, the GRU version was 2% better than the RNN version. The LSTM version achieved the best AUC score, which was improved by 3% over the GRU version.

In ref. [ 81 ], a spatiotemporal crime network (STCN) is proposed that applies a CNN for predicting crime before it occurs. The authors evaluated the STCN using 311 felony datasets from New York from 2010 to 2015. The results were extremely impressive, with the STCN achieving an F1-score of 88% and an AUC of 92%, which confirmed that it exceeded the performance of the four baselines. Their proposed model achieved the best performance in terms of both F1 and AUC, which remained better than those of the other baselines even when the time window reached 100. This study provides evidence that the system can function well even in a metropolitan area.

Proposed idea

After finding and understanding various distinct methods used by the police for surveillance purposes, we determined the importance of each method. Each surveillance method can perform well on its own and produce satisfactory results, although for only one specific characteristic, that is, if we use a Sting Ray, it can help us only when the suspect is using a phone, which should be switched on. Thus, it is only useful when the information regarding the stake out location is correct. Based on this information, we can see how the ever-evolving technology has yet again produced a smart way to conduct surveillance. The introduction of deep learning, ML, and computer vision techniques has provided us with a new perspective on ways to conduct surveillance. This is an intelligent approach to surveillance because it tries to mimic a human approach, but it does so 24 h a day, 365 days a year, and once it has been taught how to do things it does them in the same manner repeatedly.

Although we have discussed the aspects that ML and computer vision can achieve, but what are these aspects essentially? This brings us to the main point of our paper discussion, i.e., our proposed idea, which is to combine the point aspects of Sting Ray, body cams, facial recognition, number plate recognition, and stakeouts. New features iclude core analytics, neural networks, heuristic engines, recursion processors, Bayesian networks, data acquisition, cryptographic algorithms, document processors, computational linguistics, voiceprint identification, natural language processing, gait analysis, biometric recognition, pattern mining, intel interpretation, threat detection, threat classification. The new features are completely computer dependent and hence require human interaction for development; however, once developed, it functions without human interaction and frees humans for other tasks. Let us understand the use of each function.

Core analytics: This includes having knowledge of a variety of statistical techniques, and by using this knowledge, predict future outcomes, which in our case are anything from behavioral instincts to looting a store in the near future.

Neural networks: This is a concept consisting of a large number of algorithms that help in finding the relation between data by acting similar to a human brain, mimicking biological nerve cells and hence trying to think on its own, thus understanding or even predicting a crime scene.

Heuristic engines: These are engines with data regarding antiviruses, and thus knowledge about viruses, increasing the safety of our system as it identifies the type of threat and eliminates it using known antiviruses.

Cryptographic algorithms: Such algorithms are used in two parts. First, they privately encode the known confidential criminal data. Second, they are used to keep the newly discovered potential crime data encrypted.

Recursion processors: These are used to apply the functions of our machine repeatedly to make sure they continuously work and never break the surveillance of the machine.

Bayesian networks: These are probabilistic acyclic graphical models that can be used for a variety of purposes such as prediction, anomaly detection, diagnostics, automated insight, reasoning, time series prediction, and decision making under uncertainty.

Data acquisition: This might be the most important part because our system has to possess the knowledge of previous crimes and learn from them to predict future possible criminal events.

Document processors: These are used after the data collection, primarily for going through, organizing, analyzing, and learning from the data.

Computer linguistics: Using algorithms and learning models, this method is attempting to give a computer the ability to understand human spoken language, which would be ground breaking, allowing a machine to not only identify a human but also understands what the human is saying.

Natural language processor: This is also used by computers to better understand human linguistics.

Voice print identification: This is an interesting application, which tries to distinguish one person’s voice from another, making it even more recognizable and identifiable. It identifies a target with the help of certain characteristics, such as the configuration of the speaker’s mouth and throat, which can be expressed as a mathematical formula.

Gait analysis: This will be used to study human motion, understanding posture while walking. It will be used to better understand the normal pace of a person and thus judge an abnormal pace.

Bio metric identification: This is used to identify individuals by their face, or if possible, identify them by their thumb print stored in few different databases.

Pattern mining: This is a subset of data mining and helps in observing patterns among routine activities. The use of this technology will help us identify if a person is seen an usual number of times behind a pharmacy window at particular time, allowing the machine to alert the authorities.

Intel interpretation: This is also used to make sense of the information gathered, and will include almost all features mentioned above, combining the results of each and making a final meaningful prediction.

Threat detection: A threat will be detected if during the intel processing a certain number of check boxes predefined when making the system are ticked.

Threat classification: As soon as a threat is detected, it is classified, and the threat can then be categorized into criminal case levels, including burglary, murder, or a possible terrorist attack; thus, based on the time line, near or distant future threats might be predictable.

Combining all of these features, we aim to produce software that has the capability of becoming a universal police officer, having eyes and ears everywhere. Obviously, we tend to use the CCTVs in urban areas during a preliminary round to see the functioning of such software in a real-world scenario. The idea is to train and make the software learn all previously recorded crimes whose footages are available (at least 5000 cases for optimum results), through supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning to help it to understand what a crime actually is. Thus, it will achieve a better understanding of criminality and can answer how crimes happen, as well as why and where. We do not propose simply making a world-class model to predict crimes, we also suggest making it understand previous crimes to better judge and therefore better predict them.

We aim to use this type of technology on two fronts: first and most importantly, for predicting crimes before they happen, followed by a thorough analysis of a crime scene allowing the system to possibly identify aspects that even a human eye can miss.

The most interesting cutting-edge and evolutionary idea that we believe should be incorporated is the use of scenario simulations. After analyzing the scene and using the 17 main characteristics mentioned above, the software should run at least 50 simulations of the present scenario presented in front of it, which will be assisted by previously learned crime recordings. The simulation will help the software in asserting the threat level and then accordingly recommend a course of action or alert police officials.

To visualize a possible scenario where we are able to invent such software, we prepared a flow chart (Fig.  3 ) to better understand the complete process.

Flowchart of our proposed model. The data are absorbed from the surrounding with the help of cameras and microphones. If the system depicts an activity as suspicious, it gathers more intel allowing the facial algorithms to match against a big database such as a Social Security Number or Aadhaar card database. When it detects a threat, it also classifies it into categories such as the nature of the crime and time span within which it is possible to take place. With all the gathered intel and all the necessary details of the possible crime, it alerts the respective authority with a 60-word synopsis to give them a brief idea, allowing law enforcement agencies to take action accordingly

Although this paper has been implemented with high accuracy and detailed research, there are certain challenges that can pose a problem in the future. First, the correct and complete building of the whole system has to be done in the near future, allowing its implementation to take place immediately and properly. Furthermore, the implementation itself is a significant concern, as such technologies cannot be directly implemented in the open world. The system must first be tested in a small part of a metropolitan area, and only then with constant improvements (revisions of the first model) can its usage be scaled up. Hence, the challenges are more of a help in perfecting the model and thus gradually providing a perfect model that can be applied to the real world. Moreover, there are a few hurdles in the technological aspects of the model, as the size of the learning data will be enormous, and thus processing it will take days and maybe even weeks. Although these are challenges that need to be addressed, they are aspects that a collective team of experts can overcome after due diligence, and if so, the end product will be worth the hard work and persistence.

Future scope

This paper presented the techniques and methods that can be used to predict crime and help law agencies. The scope of using different methods for crime prediction and prevention can change the scenario of law enforcement agencies. Using a combination of ML and computer vision can substantially impact the overall functionality of law enforcement agencies. In the near future, by combining ML and computer vision, along with security equipment such as surveillance cameras and spotting scopes, a machine can learn the pattern of previous crimes, understand what crime actually is, and predict future crimes accurately without human intervention. A possible automation would be to create a system that can predict and anticipate the zones of crime hotspots in a city. Law enforcement agencies can be warned and prevent crime from occurring by implementing more surveillance within the prediction zone. This complete automation can overcome the drawbacks of the current system, and law enforcement agencies can depend more on these techniques in the near future. Designing a machine to anticipate and identify patterns of such crimes will be the starting point of our future study. Although the current systems have a large impact on crime prevention, this could be the next big approach and bring about a revolutionary change in the crime rate, prediction, detection, and prevention, i.e., a “universal police officer”.

Conclusions

Predicting crimes before they happen is simple to understand, but it takes a lot more than understanding the concept to make it a reality. This paper was written to assist researchers aiming to make crime prediction a reality and implement such advanced technology in real life. Although police do include the use of new technologies such as Sting Rays and facial recognition every few years, the implementation of such software can fundamentally change the way police work, in a much better way. This paper outlined a framework envisaging how the aspects of machine and deep learning, along with computer vision, can help create a system that is much more helpful to the police. Our proposed system has a collection of technologies that will perform everything from monitoring crime hotspots to recognizing people from their voice notes. The first difficulty faced will be to actually make this system, followed by problems such as its implementation and use, among others. However, all of these problems are solvable, and we can also benefit from a security system that monitors the entire city around-the-clock. In other words, to visualize a world where we incorporate such a system into a police force, tips or leads that much more reliable can be achieved and perhaps crime can be eradicated at a much faster rate.

Availability of data and materials

All relevant data and material are presented in the main paper.

Abbreviations

• Machine learning

Nondeterministic polynomial

Waikato Environment for Knowledge Analysis

K-nearest neighbor

Automatic number plate recognition

Deep neural network

Kernel density estimation

Support vector machine

Grey correlation analysis based on new weighted KNN

Autoregressive integrated moving average

Spatiotemporal crime network

Convolutional neural network

Area under the ROC curve

Recurrent neural network

Gated recurrent unit

Long short-term memory

Absolute percent error

Shah D, Dixit R, Shah A, Shah P, Shah M (2020) A comprehensive analysis regarding several breakthroughs based on computer intelligence targeting various syndromes. Augment Hum Res 5(1):14.  https://doi.org/10.1007/s41133-020-00033-z

Patel H, Prajapati D, Mahida D, Shah M (2020) Transforming petroleum downstream sector through big data: a holistic review. J Pet Explor Prod Technol 10(6):2601–2611.  https://doi.org/10.1007/s13202-020-00889-2

Szeliski R (2010) Computer vision: algorithms and applications. Springer-Verlag, Berlin, pp 1–979

MATH   Google Scholar  

Vedaldi A, Fulkerson B (2010) Vlfeat: an open and portable library of computer vision algorithms. Paper presented at the 18th ACM international conference on multimedia. ACM, Firenze. https://doi.org/10.1145/1873951.1874249

Le TL, Nguyen MQ, Nguyen TTM (2013) Human posture recognition using human skeleton provided by Kinect. In: Paper presented at the 2013 international conference on computing, management and telecommunications. IEEE, Ho Chi Minh City. https://doi.org/10.1109/ComManTel.2013.6482417

Ahir K, Govani K, Gajera R, Shah M (2020) Application on virtual reality for enhanced education learning, military training and sports. Augment Hum Res 5(1):7. ( https://doi.org/10.1007/s41133-019-0025-2 )

Talaviya T, Shah D, Patel N, Yagnik H, Shah M (2020) Implementation of artificial intelligence in agriculture for optimisation of irrigation and application of pesticides and herbicides. Artif Intell Agric 4:58–73.  https://doi.org/10.1016/j.aiia.2020.04.002

Jha K, Doshi A, Patel P, Shah M (2019) A comprehensive review on automation in agriculture using artificial intelligence. Artif Intell Agric 2:1–12.  https://doi.org/10.1016/j.aiia.2019.05.004

Kakkad V, Patel M, Shah M (2019) Biometric authentication and image encryption for image security in cloud framework. Multiscale Multidiscip Model Exp Des 2(4):233–248.  https://doi.org/10.1007/s41939-019-00049-y

Pathan M, Patel N, Yagnik H, Shah M (2020) Artificial cognition for applications in smart agriculture: a comprehensive review. Artif Intell Agric 4:81–95.  https://doi.org/10.1016/j.aiia.2020.06.001

Pandya R, Nadiadwala S, Shah R, Shah M (2020) Buildout of methodology for meticulous diagnosis of K-complex in EEG for aiding the detection of Alzheimer's by artificial intelligence. Augment Hum Res 5(1):3.  https://doi.org/10.1007/s41133-019-0021-6

Dey A (2016) Machine learning algorithms: a review. Int J Comput Sci Inf Technol 7(3):1174–1179

Google Scholar  

Sukhadia A, Upadhyay K, Gundeti M, Shah S, Shah M (2020) Optimization of smart traffic governance system using artificial intelligence. Augment Hum Res 5(1):13.  https://doi.org/10.1007/s41133-020-00035-x

Musumeci F, Rottondi C, Nag A, Macaluso I, Zibar D, Ruffini M et al (2019) An overview on application of machine learning techniques in optical networks. IEEE Commun Surv Tutorials 21(2):1381–1408.  https://doi.org/10.1109/COMST.2018.2880039

Patel D, Shah Y, Thakkar N, Shah K, Shah M (2020) Implementation of artificial intelligence techniques for cancer detection. Augment Hum Res 5(1):6. https://doi.org/10.1007/s41133-019-0024-3

Kundalia K, Patel Y, Shah M (2020) Multi-label movie genre detection from a movie poster using knowledge transfer learning. Augment Hum Res 5(1):11. https://doi.org/10.1007/s41133-019-0029-y

Article   Google Scholar  

Marsland S (2015) Machine learning: an algorithmic perspective. CRC Press, Boca Raton, pp 1–452. https://doi.org/10.1201/b17476-1

Jani K, Chaudhuri M, Patel H, Shah M (2020) Machine learning in films: an approach towards automation in film censoring. J Data Inf Manag 2(1):55–64. https://doi.org/10.1007/s42488-019-00016-9

Parekh V, Shah D, Shah M (2020) Fatigue detection using artificial intelligence framework. Augment Hum Res 5(1):5 https://doi.org/10.1007/s41133-019-0023-4

Gandhi M, Kamdar J, Shah M (2020) Preprocessing of non-symmetrical images for edge detection. Augment Hum Res 5(1):10 https://doi.org/10.1007/s41133-019-0030-5

Panchiwala S, Shah M (2020) A comprehensive study on critical security issues and challenges of the IoT world. J Data Inf Manag 2(7):257–278. https://doi.org/10.1007/s42488-020-00030-2

Simon A, Deo MS, Venkatesan S, Babu DR (2016) An overview of machine learning and its applications. Int J Electr Sci Eng 1(1):22–24.

Parekh P, Patel S, Patel N, Shah M (2020) Systematic review and meta-analysis of augmented reality in medicine, retail, and games. Vis Comput Ind Biomed Art 3(1):21. https://doi.org/10.1186/s42492-020-00057-7

Shah K, Patel H, Sanghvi D, Shah M (2020) A comparative analysis of logistic regression, random forest and KNN models for the text classification. Augment Hum Res 5(1):12. https://doi.org/10.1007/s41133-020-00032-0

Patel D, Shah D, Shah M (2020) The intertwine of brain and body: a quantitative analysis on how big data influences the system of sports. Ann Data Sci 7(1):1–16. https://doi.org/10.1007/s40745-019-00239-y

Judd S (1988) On the complexity of loading shallow neural networks. J Complex 4(3):177–192. https://doi.org/10.1016/0885-064X(88)90019-2

Article   MathSciNet   MATH   Google Scholar  

Blum AL, Rivest RL (1992) Training a 3-node neural network is NP-complete. Neural Netw 5(1):117–127. https://doi.org/10.1016/S0893-6080(05)80010-3

Gupta A, Dengre V, Kheruwala HA, Shah M (2020) Comprehensive review of text-mining applications in finance. Financ Innov 6(1):1–25. https://doi.org/10.1186/s40854-020-00205-1

Shah N, Engineer S, Bhagat N, Chauhan H, Shah M (2020) Research trends on the usage of machine learning and artificial intelligence in advertising. Augment Hum Res 5(1):19. https://doi.org/10.1007/s41133-020-00038-8

Naik B, Mehta A, Shah M (2020) Denouements of machine learning and multimodal diagnostic classification of Alzheimer's disease. Vis Comput Ind Biomed Art 3(1):26. https://doi.org/10.1186/s42492-020-00062-w

Chen P, Yuan HY, Shu XM (2008) Forecasting crime using the ARIMA model. In: Paper presented at the 5th international conference on fuzzy systems and knowledge discovery. IEEE, Ji'nan 18-20 October 2008. https://doi.org/10.1109/FSKD.2008.222

Rani A, Rajasree S (2014) Crime trend analysis and prediction using mahanolobis distance and dynamic time warping technique. Int J Comput Sci Inf Technol 5(3):4131–4135

Gorr W, Harries R (2003) Introduction to crime forecasting. Int J Forecast 19(4):551–555. https://doi.org/10.1016/S0169-2070(03)00089-X

Rummens A, Hardyns W, Pauwels L (2017) The use of predictive analysis in spatiotemporal crime forecasting: building and testing a model in an urban context. Appl Geogr 86:255–261. https://doi.org/10.1016/j.apgeog.2017.06.011

Bates A (2017) Stingray: a new frontier in police surveillance. Cato Institute Policy Analysis, No. 809

Joh EE (2017) The undue influence of surveillance technology companies on policing. N Y Univ Law Rev 92:101–130. https://doi.org/10.2139/ssrn.2924620

Vredeveldt A, Kesteloo L, Van Koppen PJ (2018) Writing alone or together: police officers' collaborative reports of an incident. Crim Justice Behav 45(7):1071–1092. https://doi.org/10.1177/0093854818771721

McNeal GS (2014) Drones and aerial surveillance: considerations for legislators. In: Brookings Institution: The Robots Are Coming: The Project On Civilian Robotics, November 2014, Pepperdine University Legal Studies Research Paper No. 2015/3

Fatih T, Bekir C (2015) Police use of technology to fight against crime. Eur Sci J 11(10):286–296

Katz CM, Choate DE, Ready JR, Nuňo L (2014) Evaluating the impact of officer worn body cameras in the Phoenix Police Department. Center for Violence Prevention & Community Safety, Arizona State University, Phoenix, pp 1–43

Stanley J (2015) Police body-mounted cameras: with right policies in place, a win for all. https://www.aclu.org/police-body-mounted-cameras-right-policies-place-win-all . Accessed 15 Aug 2015

McClendon L, Meghanathan N (2015) Using machine learning algorithms to analyze crime data. Mach Lear Appl Int J 2(1):1–12. https://doi.org/10.5121/mlaij.2015.2101

Frank E, Hall M, Trigg L, Holmes G, Witten IH (2004) Data mining in bioinformatics using Weka. Bioinformatics 20(15):2479–2481. https://doi.org/10.1093/bioinformatics/bth261

Kim S, Joshi P, Kalsi PS, Taheri P (2018) Crime analysis through machine learning. In: Paper presented at the IEEE 9th annual information technology, electronics and mobile communication conference. IEEE, Vancouver 1-3 November 2018. https://doi.org/10.1109/IEMCON.2018.8614828

Tabedzki C, Thirumalaiswamy A, van Vliet P (2018) Yo home to Bel-Air: predicting crime on the streets of Philadelphia. In: University of Pennsylvania, CIS 520: machine learning

Bharati A, Sarvanaguru RAK (2018) Crime prediction and analysis using machine learning. Int Res J Eng Technol 5(9):1037–1042

Prithi S, Aravindan S, Anusuya E, Kumar AM (2020) GUI based prediction of crime rate using machine learning approach. Int J Comput Sci Mob Comput 9(3):221–229

Kang HW, Kang HB (2017) Prediction of crime occurrence from multi-modal data using deep learning. PLoS One 12(4):e0176244. https://doi.org/10.1371/journal.pone.0176244

Bandekar SR, Vijayalakshmi C (2020) Design and analysis of machine learning algorithms for the reduction of crime rates in India. Procedia Comput Sci 172:122–127. https://doi.org/10.1016/j.procs.2020.05.018

Hossain S, Abtahee A, Kashem I, Hoque M, Sarker IH (2020) Crime prediction using spatio-temporal data. arXiv preprint arXiv:2003.09322. https://doi.org/10.1007/978-981-15-6648-6_22

Stalidis P, Semertzidis T, Daras P (2018) Examining deep learning architectures for crime classification and prediction. arXiv preprint arXiv: 1812.00602. p. 1–13

Jha P, Jha R, Sharma A (2019) Behavior analysis and crime prediction using big data and machine learning. Int J Recent Technol Eng 8(1):461–468

Tyagi D, Sharma S (2018) An approach to crime data analysis: a systematic review. Int J Eng Technol Manag Res 5(2):67–74. https://doi.org/10.29121/ijetmr.v5.i2.2018.615

Lin YL, Yen MF, Yu LC (2018) Grid-based crime prediction using geographical features. ISPRS Int J Geo-Inf 7(8):298. https://doi.org/10.3390/ijgi7080298

Ahishakiye E, Taremwa D, Omulo EO, Niyonzima I (2017) Crime prediction using decision tree (J48) classification algorithm. Int J Comput Inf Technol 6(3):188–195

Sun CC, Yao CL, Li X, Lee K (2014) Detecting crime types using classification algorithms. J Digit Inf Manag 12(8):321–327. https://doi.org/10.14400/JDC.2014.12.8.321

Shojaee S, Mustapha A, Sidi F, Jabar MA (2013) A study on classification learning algorithms to predict crime status. Int J Digital Content Technol Appl 7(9):361–369

Obuandike GN, Isah A, Alhasan J (2015) Analytical study of some selected classification algorithms in WEKA using real crime data. Int J Adv Res Artif Intell 4(12):44–48. https://doi.org/10.14569/IJARAI.2015.041207

Iqbal R, Murad MAA, Mustapha A, Panahy PHS, Khanahmadliravi N (2013) An experimental study of classification algorithms for crime prediction. Indian J Sci Technol 6(3):4219–4225. https://doi.org/10.17485/ijst/2013/v6i3.6

Jangra M, Kalsi S (2019) Crime analysis for multistate network using naive Bayes classifier. Int J Comput Sci Mob Comput 8(6):134–143

Wibowo AH, Oesman TI (2020) The comparative analysis on the accuracy of k-NN, naive Bayes, and decision tree algorithms in predicting crimes and criminal actions in Sleman regency. J Phys Conf Ser 1450:012076. https://doi.org/10.1088/1742-6596/1450/1/012076

Vanhoenshoven F, Nápoles G, Bielen S, Vanhoof K (2017) Fuzzy cognitive maps employing ARIMA components for time series forecasting. In: Czarnowski I, Howlett RJ, Jain LC (eds) Proceedings of the 9th KES international conference on intelligent decision technologies 2017, vol 72. Springer, Heidelberg, pp 255–264. https://doi.org/10.1007/978-3-319-59421-7_24

Chapter   Google Scholar  

Gorr W, Olligschlaeger AM, Thompson Y (2000) Assessment of crime forecasting accuracy for deployment of police. Int J Forecast 2000:743–754

Yu CH, Ward MW, Morabito M, Ding W (2011) Crime forecasting using data mining techniques. In: Paper presented at the 2011 IEEE 11th international conference on data mining workshops. IEEE, Vancouver 11-11 December 2011. https://doi.org/10.1109/ICDMW.2011.56

Alves LGA, Ribeiro HV, Rodrigues FA (2018) Crime prediction through urban metrics and statistical learning. Phys A Stat Mech Appl 505:435–443. https://doi.org/10.1016/j.physa.2018.03.084

Idrees H, Shah M, Surette R (2018) Enhancing camera surveillance using computer vision: a research note. Polic Int J 41(2):292–307. https://doi.org/10.1108/PIJPSM-11-2016-0158

Wu G, Wu Y, Jiao L, Wang YF, Chang EY (2003) Multi-camera spatio-temporal fusion and biased sequence-data learning for security surveillance. In: Paper presented at the 11th ACM international conference on multimedia. ACM, Berkeley 2-8 November 2003. https://doi.org/10.1145/957013.957126

Wang YF, Chang EY, Cheng KP (2005) A video analysis framework for soft biometry security surveillance. In: Paper presented at the 3rd ACM international workshop on video surveillance & sensor networks. ACM, Hilton 11 November 2005. https://doi.org/10.1145/1099396.1099412

Shah M, Javed O, Shafique K (2007) Automated visual surveillance in realistic scenarios. IEEE MultiMed 14(1):30–39. https://doi.org/10.1109/MMUL.2007.3

Burton AM, Wilson S, Cowan M, Bruce V (1999) Face recognition in poor-quality video: evidence from security surveillance. Psychol Sci 10(3):243–248. https://doi.org/10.1111/1467-9280.00144

Goyal A, Bhatia R (2016) Automated car number plate detection system to detect far number plates. IOSR J Comput Eng 18(4):34–40. https://doi.org/10.9790/0661-1804033440

Wang B, Yin PH, Bertozzi AL, Brantingham PJ, Osher SJ, Xin J (2019) Deep learning for real-time crime forecasting and its ternarization. Chin Ann Math Ser B 40(6):949–966. https://doi.org/10.1007/s11401-019-0168-y

Stec A, Klabjan D (2018) Forecasting crime with deep learning. arXiv preprint arXiv:1806.01486. p. 1–20

Chen YL, Wu BF, Huang HY, Fan CJ (2011) A real-time vision system for nighttime vehicle detection and traffic surveillance. IEEE Trans Ind Electron 58(5):2030–2044. https://doi.org/10.1109/TIE.2010.2055771

Poppe R (2007) Vision-based human motion analysis: an overview. Comput Vision Image Underst 108(1–2):4–18. https://doi.org/10.1016/j.cviu.2006.10.016

Najjar A, Kaneko S, Miyanaga Y (2018) Crime mapping from satellite imagery via deep learning. arXiv preprint arXiv:1812.06764. p. 1–8

Khosla A, An B, Lim JJ, Torralba A (2014) Looking beyond the visible scene. In: Paper presented at the of IEEE conference on computer vision and pattern recognition. IEEE, Columbus 23-28 June 2014. https://doi.org/10.1109/CVPR.2014.474

Dee HM, Velastin SA (2008) How close are we to solving the problem of automated visual surveillance? Mach Vis Appl 19(5–6):329–343. https://doi.org/10.1007/s00138-007-0077-z

Rautaray SS (2012) Real time hand gesture recognition system for dynamic applications. Int J Ubi Comp 3(1):21–31. https://doi.org/10.5121/iju.2012.3103

Zhuang Y, Almeida M, Morabito M, Ding W (2017) Crime hot spot forecasting: a recurrent model with spatial and temporal information. In: Paper presented at the IEEE international conference on big knowledge. IEEE, Hefei 9-10 August 2017. https://doi.org/10.1109/ICBK.2017.3

Duan L, Hu T, Cheng E, Zhu JF, Gao C (2017) Deep convolutional neural networks for spatiotemporal crime prediction. In: Paper presented at the 16th international conference information and knowledge engineering. CSREA Press, Las Vegas 17-20 July 2017

Download references

Acknowledgements

The authors are grateful to Department of Computer Engineering, SAL Institute of Technology and Engineering Research and Department of Chemical Engineering, School of Technology, Pandit Deendayal Energy University for the permission to publish this research.

Not applicable.

Author information

Authors and affiliations.

Department of Computer Engineering, Sal Institute of Technology and Engineering Research, Ahmedabad, Gujarat, 380060, India

Neil Shah & Nandish Bhagat

Department of Chemical Engineering, School of Technology, Pandit Deendayal Energy University, Gandhinagar, Gujarat, 382426, India

You can also search for this author in PubMed   Google Scholar

Contributions

All the authors make substantial contribution in this manuscript; NS, NB and MS participated in drafting the manuscript; NS, NB and MS wrote the main manuscript; all the authors discussed the results and implication on the manuscript at all stages the author(s) read and approved the final manuscript.

Corresponding author

Correspondence to Manan Shah .

Ethics declarations

Competing interests.

The authors declare that they have no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Shah, N., Bhagat, N. & Shah, M. Crime forecasting: a machine learning and computer vision approach to crime prediction and prevention. Vis. Comput. Ind. Biomed. Art 4 , 9 (2021). https://doi.org/10.1186/s42492-021-00075-z

Download citation

Received : 18 July 2020

Accepted : 05 April 2021

Published : 29 April 2021

DOI : https://doi.org/10.1186/s42492-021-00075-z

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Computer vision
• Crime forecasting

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Published: 30 June 2022

Event-level prediction of urban crime reveals a signature of enforcement bias in US cities

• Victor Rotaru 1 , 2 ,
• Yi Huang 1 ,
• Timmy Li 1 , 2 ,
• James Evans   ORCID: orcid.org/0000-0001-9838-0707 3 , 4 , 5 &
• Ishanu Chattopadhyay   ORCID: orcid.org/0000-0001-8339-8162 1 , 4 , 6  

Nature Human Behaviour volume  6 ,  pages 1056–1068 ( 2022 ) Cite this article

5474 Accesses

9 Citations

1014 Altmetric

Metrics details

• Complex networks
• Criminology

Policing efforts to thwart crime typically rely on criminal infraction reports, which implicitly manifest a complex relationship between crime, policing and society. As a result, crime prediction and predictive policing have stirred controversy, with the latest artificial intelligence-based algorithms producing limited insight into the social system of crime. Here we show that, while predictive models may enhance state power through criminal surveillance, they also enable surveillance of the state by tracing systemic biases in crime enforcement. We introduce a stochastic inference algorithm that forecasts crime by learning spatio-temporal dependencies from event reports, with a mean area under the receiver operating characteristic curve of ~90% in Chicago for crimes predicted per week within ~1,000 ft. Such predictions enable us to study perturbations of crime patterns that suggest that the response to increased crime is biased by neighbourhood socio-economic status, draining policy resources from socio-economically disadvantaged areas, as demonstrated in eight major US cities.

This is a preview of subscription content, access via your institution

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 12 digital issues and online access to articles

111,21 € per year

only 9,27 € per issue

Buy this article

• Purchase on SpringerLink
• Instant access to full article PDF

Prices may be subject to local taxes which are calculated during checkout

Similar content being viewed by others

Socio-economic, built environment, and mobility conditions associated with crime: a study of multiple cities

A multisource database tracking the impact of the COVID-19 pandemic on the communities of Boston, MA, USA

Using machine learning to forecast domestic homicide via police data and super learning

Data availability.

Crime incident data used in this study are in the public domain. The web links for the data sources for seven out of the eight cities considered here are: opendata.atlantapd.org , data.austintexas.gov , data.detroitmi.gov , data.lacity.org , www.opendata.philly.org , data.sfgov.org , and data.cityofchicago.org , and for Portland the data along with the leader-board data for the forecasting challenge were obtained from nij.ojp.gov .

Code availability

Software with source code is available at https://github.com/zeroknowledgediscovery/Cynet , and the current version of the software may be referenced by https://doi.org/10.5281/zenodo.5730613 . Any questions on implementation should be directed to the corresponding author.

Bowers, K. J., Johnson, S. D. & Pease, K. Prospective hot-spotting: the future of crime mapping? Br. J. Criminol. 44 , 641–658 (2004).

Article   Google Scholar  

Chainey, S., Tompson, L. & Uhlig, S. The utility of hotspot mapping for predicting spatial patterns of crime. Secur. J. 21 , 4–28 (2008).

Fielding, M. & Jones, V. ‘Disrupting the optimal forager’: predictive risk mapping and domestic burglary reduction in Trafford, Greater Manchester. Int. J. Police Sci. Manage. 14 , 30–41 (2012).

Mohler, G. O., Short, M. B., Brantingham, P. J., Schoenberg, F. P. & Tita, G. E. Self-exciting point process modeling of crime. J. Am. Stat. Assoc. 106 , 100–108 (2011).

Article   CAS   Google Scholar  

Mohler, G. O. et al. Randomized controlled field trials of predictive policing. J. Am. Stat. Assoc. 110 , 1399–1411 (2015).

Poisson, S. D. Probabilité des Jugements en Matiére Criminelle et en Matiére Civile, Précédées des Régles Générales du Calcul des Probabilitiés (Bachelier, 1837).

Du Sautoy, M. The Creativity Code: Art and Innovation in the Age of AI (Harvard Univ. Press, 2020).

Ferdinand, T. N. Demographic shifts and criminality: an inquiry. Br. J. Criminol. 10 , 169–175 (1970).

Cohen, L. & Felson, M. Social change and crime rate trends: a routine activity approach. Am. Sociol. Rev. 44 , 588–608 (1979).

Cohen, L. E. Modeling crime trends: a criminal opportunity perspective. J. Res. Crime Delinquency 18 , 138–164 (1981).

Wang, X. & Brown, D. E. The spatio-temporal modeling for criminal incidents. Secur. Inform. 1 , 2 (2012).

Liu, H. & Brown, D. E. Criminal incident prediction using a point-pattern-based density model. 19 , 603–622 (2003).

Caplan, J. M., Kennedy, L. W., Barnum, J. D. & Piza, E. L. Crime in context: utilizing risk terrain modeling and conjunctive analysis of case configurations to explore the dynamics of criminogenic behavior settings. J. Contemp. Crim. Justice 33 , 133–151 (2017).

Kang, H. W. & Kang, H. B. Prediction of crime occurrence from multi-modal data using deep learning. PLoS ONE 12 , e0176244 (2017).

Article   PubMed   PubMed Central   CAS   Google Scholar  

Flaxman, S., Chirico, M., Pereira, P. & Loeffler, C. Scalable high-resolution forecasting of sparse spatiotemporal events with kernel methods: a winning solution to the NIJ ‘real-time crime forecasting challenge’. Ann. Appl. Stat. 13 , 2564–2585 (2019).

Mohler, G. & Porter, M. D. Rotational grid, PAI-maximizing crime forecasts. Stat. Anal. Data Min. 11 , 227–236 (2018).

Chattopadhyay, I. & Lipson, H. Abductive learning of quantized stochastic processes with probabilistic finite automata. Philos. Trans. R. Soc. A 371 , 20110543 (2013).

Mohri, M. Weighted Finite-State Transducer Algorithms. An Overview (Springer, 2004).

Granger, C. W. J. Testing for causality: a personal viewpoint. J. Econ. Dyn. Control 2 , 329 – 352 (1980).

Papachristos, A. V. & Bastomski, S. Connected in crime: the enduring effect of neighborhood networks on the spatial patterning of violence. Am. J. Sociol. 124 , 517–568 (2018).

Papachristos, A. V., Wildeman, C. & Roberto, E. Tragic, but not random: the social contagion of nonfatal gunshot injuries. Soc. Sci. Med. 125 , 139–150 (2015).

Article   PubMed   Google Scholar  

Green, B., Horel, T. & Papachristos, A. V. Modeling contagion through social networks to explain and predict gunshot violence in Chicago, 2006 to 2014. JAMA Intern. Med. 177 , 326–333 (2017).

Kang, H.-W. & Kang, H.-B. Prediction of crime occurrence from multi-modal data using deep learning. PLoS ONE 12 , e0176244 (2017).

Stec, A. & Klabjan, D. Forecasting crime with deep learning. Preprint at https://arxiv.org/abs/1806.01486 (2018).

Hannon, L. Neighborhood residence and assessments of racial profiling using census data. Socius 5 , 2378023118818746 (2019).

Meyer, W. B. & Graybill, J. K. The suburban bias of American society? Urban Geogr. 37 , 863–882 (2016).

Lipton, M. et al. Why Poor People Stay Poor: a Study of Urban Bias in World Development (Australian National Univ. Press, 1977).

Sternlieb, G. & Jackson, K. T. Crabgrass frontier: the suburbanization of the United States. Political Sci. Q. 101 , 493 (1986).

Duany, A., Plater-Zyberk, E. & Speck, J. Suburban nation: the rise of sprawl and the decline of the American dream. Choice Rev. Online 38 , 38–1251–38–1251 (2000).

Google Scholar  

Lazare, D. America’s Undeclared War: What’s Killing Our Cities and How to Stop It (Harcourt, 2001).

Young, I. M. Inclusion and Democracy (Oxford Univ. Press, 2002).

Kaplan, M. S., Crespo, C. J., Huguet, N. & Marks, G. Ethnic/racial homogeneity and sexually transmitted disease: a study of 77 Chicago community areas. Sex. Transm. Dis. 36 , 108–111 (2009).

Sherman, L. W., Gartin, P. R. & Buerger, M. E. Hot spots of predatory crime: routine activities and the criminology of place. Criminology 27 , 27–56 (1989).

Wooldredge, J. Examining the (ir)relevance of aggregation bias for multilevel studies of neighborhoods and crime with an example comparing census tracts to official neighborhoods in Cincinnati. Criminology 40 , 681–710 (2002).

Mears, D. P. & Bhati, A. S. No community is an island: the effects of resource deprivation on urban violence in spatially and socially proximate communities. Criminology 44 , 509–548 (2006).

Weisburd, D., Groff, E. R., Yang, S.-M. & Telep, C. W. Criminology of Place (Springer, 2014).

Small, M. L. Four reasons to abandon the idea of ‘the ghetto’. City Community 7 , 389–398 (2008).

Baumgarten, M. Ghetto: the invention of a place, the history of an idea. Jew. Q. 63 , 62–63 (2016).

Heaven, W. D. Predictive policing algorithms are racist. They need to be dismantled. MIT ZTechnol. Rev. 17 , 2020 (2020).

Brayne, S. & Christin, A. Technologies of crime prediction: the reception of algorithms in policing and criminal courts. Social Problems 68 , 608–624 (2020).

St. Louis, S. & Greene, J. R. Social context in police legitimacy: giving meaning to police/community contacts. Policing Soc. 30 , 656–673 (2020).

Weisburd, D. Place-based policing. Ideas in American Policing 9 , 1–16 (2008).

Kushnick, L. ‘Over policed and under protected’: Stephen lawrence, institutional and police practices. Sociol. Res. Online 4 , 156–166 (1999).

Clifford, R. S. Juvenile delinquency and urban areas: a study of rates of delinquents in relation to differential characteristics of local communities in American cities. Am. J. Sociol. 49 , 100–101 (1943).

Sampson, R. J., Raudenbush, S. W. & Earls, F. Neighborhoods and violent crime: a multilevel study of collective efficacy. Science 277 , 918–924 (1997).

Article   CAS   PubMed   Google Scholar  

Miethe, T. D., Hughes, M. & McDowall, D. Social change and crime rates: an evaluation of alternative theoretical approaches. Soc. Forces 70 , 165–185 (1991).

Braga, A. A. & Clarke, R. V. Explaining high-risk concentrations of crime in the city: social disorganization, crime opportunities, and important next steps. J. Rs. Crime Delinquency 51 , 480–498 (2014).

Silver, D. & Clark, T. Scenescapes: How Qualities of Place Shape Social Life (Univ. of Chicago Press, 2016).

Nathan, R. P. & Adams, C. F. Four perspectives on urban hardship. Political Sci. Q. 104 , 483–508 (1989).

Granger, C. W. J. Testing for causality. J. Econ. Dyn. Control 2 , 329–352 (1980).

Montero-Manso, P. & Hyndman, R. J. Principles and algorithms for forecasting groups of time series: locality and globality. Int. J. Forecast. 37 , 1632–1653 (2021).

Seabold, S. & Perktold, J. Statsmodels: econometric and statistical modeling with Python. In Proc. 9th Python in Science Conference https://conference.scipy.org/proceedings/scipy2010/pdfs/seabold.pdf (2010).

Laxy, M., Malecki, K. C., Givens, M. L., Walsh, M. C. & Nieto, F. J. The association between neighborhood economic hardship, the retail food environment, fast food intake, and obesity: findings from the Survey of the Health of Wisconsin. BMC Public Health 15 , 1–10 (2015).

Download references

Acknowledgements

Our work greatly benefited from discussion of everyone who participated in our workshop series on crime prediction at the Neubauer Collegium for culture and society ( https://neubauercollegium.uchicago.edu/events/uc/crimes_of_prediction_workshop/ ), and with those with whom we had extended conversations to ground and refine our modelling approach.

Data were provided by the City of Chicago data portal at https://data.cityofchicago.org . The City of Chicago (‘City’) voluntarily provides the data on this website as a service to the public. The City makes no warranty, representation, or guarantee as to the content, accuracy, timeliness, or completeness of any of the data provided at this website ( https://www.chicago.gov/city/en/narr/foia/data_disclaimer.html ), and the authors of this study are solely responsible for the opinions and conclusions expressed in this study. Sources of the crime incidence data for the other cities are tabulated in Table 1 . Socio-economic data for metropolitan areas were obtained from https://www.census.gov .

This work is funded in part by the Defense Sciences Office of the Defense Advanced Research Projects Agency projects HR00111890043/P00004 and W911NF2010302, and the Neubauer Collegium for Culture and Society through the Faculty Initiated Research Program 2017. The claims made in this study do not necessarily reflect the position or the policy of the sponsors, and no official endorsement should be inferred.

Author information

Authors and affiliations.

Department of Medicine, University of Chicago, Chicago, IL, USA

Victor Rotaru, Yi Huang, Timmy Li & Ishanu Chattopadhyay

Department of Computer Science, University of Chicago, Chicago, IL, USA

Victor Rotaru & Timmy Li

Department of Sociology, University of Chicago, Chicago, IL, USA

James Evans

Committee on Quantitative Methods in Social, Behavioral, and Health Sciences, University of Chicago, Chicago, IL, USA

James Evans & Ishanu Chattopadhyay

Santa Fe Institute, Santa Fe, NM, USA

Committee on Genetics, Genomics, and Systems Biology, University of Chicago, Chicago, IL, USA

Ishanu Chattopadhyay

You can also search for this author in PubMed   Google Scholar

Contributions

Y.H. and I.C. worked out key mathematical details of the Granger network framework. V.R., T.L., Y.H. and I.C. implemented the algorithms and generated results. Y.H., V.R. and T.L. contributed equally in realizing the current implementation of the software. I.C. generated the visualizations in this study. J.E. provided key insights into modelling and interpreting social dynamics. J.E. and I.C. conceived and designed the research, and wrote the paper.

Corresponding author

Correspondence to Ishanu Chattopadhyay .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Nature Human Behaviour thanks the anonymous reviewers for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended data fig. 1 out of sample predictive performance over the years..

We show that the predictive performance is very stable, and variation in mean AUC is limited to the third place of decimal, at least when analyzing the last few years (4 years shown).

Extended Data Fig. 2 Comparison of Predicted vs Actual Sample Paths in Time and Frequency Domains.

Panels a, c and e show that the predicted and actual sample paths are pretty close for different years, when compared over the first 150 days of each year. Panels b, d and f show that the Fourier coefficients match up pretty well as well. More importantly, while our models do not explicitly incorporate any periodic elements that are being tuned, we still manage to capture the weekly, (approximately) biweekly and longer periodic regularities.

Extended Data Fig. 3 Perturbation Effects Across Variables.

We see that the decrease of violent crimes from increase of property crimes are localized in disadvantaged neighborhoods (panel g). Similarly, the decrease of property crimes from increase of violent crimes is also localized to disadvantaged neighborhoods (panel a), as well as the decreased violent crimes from increased arrests (panel k). We see a weaker localization for the corresponding increases in crime rates under similar perturbations. Looking at other pairs of variables under perturbation (rest of the panels), we generally do not see a very prominent correspondence with the distribution of socio-economic indicators. It seems crimes (and particularly violent crimes) are easier to dampen in locales with high existing crime rates, which is desirable result. But such conclusions are currently confounded by SES variables, and further work is needed to investigate these effects more thoroughly.

Extended Data Fig. 4 Stability of Suburban Bias over Years (Violent Crimes).

We show that the nature of the perturbation response shown in Fig. 3 holds true for earlier years as well: panels a and b correspond to year 2014, c and d correspond to 2015 and e and f correspond to year 2016, all of which follow the same pattern shown in Fig. 3 .

Extended Data Fig. 5 Stability of Suburban Bias over Years (Property Crimes).

Extended data fig. 6 automatic neighborhood decomposition using event predictability..

Using Event Predictability Computing a bi-clustering on the source-vs-target influence matrix (panel A) isolates a set of spatial tiles that are, on average, good predictors for all other tiles. Using this set, we use a Voronoi decomposition of the city (Panel B), which realizes an automatic spatial decomposition of the urban space, driven by event predictability.

Supplementary information

Supplementary Fig. 1 and methods.

Reporting Summary

Rights and permissions.

Reprints and permissions

About this article

Cite this article.

Rotaru, V., Huang, Y., Li, T. et al. Event-level prediction of urban crime reveals a signature of enforcement bias in US cities. Nat Hum Behav 6 , 1056–1068 (2022). https://doi.org/10.1038/s41562-022-01372-0

Download citation

Received : 30 January 2021

Accepted : 02 May 2022

Published : 30 June 2022

Issue Date : August 2022

DOI : https://doi.org/10.1038/s41562-022-01372-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

The promises and perils of crime prediction.

• Andrew V. Papachristos

Nature Human Behaviour (2022)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

• Systematic Review
• Open access
• Published: 27 May 2020

A systematic review on spatial crime forecasting

• Ourania Kounadi 1 ,
• Alina Ristea   ORCID: orcid.org/0000-0003-2682-1416 2 , 3 ,
• Adelson Araujo Jr. 4 &
• Michael Leitner 2 , 5  

Crime Science volume  9 , Article number:  7 ( 2020 ) Cite this article

18k Accesses

50 Citations

9 Altmetric

Metrics details

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.

Limitations

Current studies often lack a clear reporting of study experiments, feature engineering procedures, and are using inconsistent terminology to address similar problems.

Conclusions

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.

Implications

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.

Environmental criminology provides an important theoretical foundation for exploring and understanding spatial crime distribution (Bruinsma and Johnson 2018 ). The occurrence of crime within an area fluctuates from place to place. Besides, crime occurrences depend on a multitude of factors, and they show an increased strategic complexity and interaction with other networks, such as institutional or community-based. In criminology research, these factors are primarily referred to as crime attractors and crime generators (Kinney et al. 2008 ). Spatial fluctuations and dependencies to attractors and generators suggest that crime is not random in time and in space. A strong foundation for spatial predictive crime analytics is the Crime Pattern Theory (Brantingham and Brantingham 1984 ). It is used to explain why crimes occur in specific areas, suggests that crime is not random, and that it can be organized or opportunistic. In particular, it shows that when the activity space of a victim intersects with the activity space of an offender, there are higher chances for a crime occurrence. The activity perimeter of a person is spatially constrained by locations that are attended (nodes). For example, if one of the personal nodes is in a high-crime neighbourhood, criminals come across new opportunities to offend.

If crime is not random it can be studied, and as such, its patterns, including the spatial component, can be modelled. As a consequence, environmental criminology theories have been tested scientifically and in the past decade various research fields have made much progress in developing methods for (spatiotemporal) crime prediction and evaluation (Caplan et al. 2011 ; Mohler et al. 2011 , 2015 ; Perry 2013 ; Wang and Brown 2011 ; Yu et al. 2011 ).

Most prediction techniques are used for retrospective forecasting, i.e., predicting the future through historical data. Historical crime data are used alone or together with crime attractors and generators (which can be demographic, environmental, etc.) in diverse types of prediction models (Mohler et al. 2011 ; Ohyama and Amemiya 2018 ; Yu et al. 2011 ). Apart from static data, such as demographics or socio-economic variables, as predictors, researchers have recently included dynamic space and time features, thus giving a boost to predicting crime occurrences. These models consist of social media data (Al Boni and Gerber 2016 ; Gerber 2014 ; Kadar et al. 2017 ; Wang et al. 2012 ; Williams and Burnap 2015 ), and taxi pick-up and drop-off data (Kadar and Pletikosa 2018 ; Wang et al. 2016 ; Zhao and Tang 2017 ).

Although current crime prediction models show increasing accuracy, little emphasis has been placed on drawing the empirical and technical landscape to outline strengths and opportunities for future research, but also to identify weaknesses and threats. In this paper, we focus on spatial crime forecasting, which is the spatial forecasting of crime-related information. It has many applications such as the spatial forecast of the number of crimes, the type of criminal activity, the next location of a crime in a series, or other crime-related information. At this point, we should note that we came across papers that claim to do spatial crime forecasting or crime forecasting while extrapolating in space or detecting spatial clusters. Overall, papers in the field of spatial crime analysis use the term prediction synonymous with forecasting and they have a preference for the term prediction (Perry 2013 ). However, there are several spatial prediction types with applications of interpolation or extrapolation. Forecasting is a prediction that extrapolates an estimated variable into a future time. While prediction can be synonymous with forecasting, it is often also used to infer unknown values regardless of the time dimension (e.g., predict the crime in area A using a model derived from area B). Cressie ( 1993 , pp 105–106) refers to spatial prediction as an inference process to predict values at unknown locations from data observed at known locations. His terminology includes the temporal notions of smoothing or interpolation, filtering, and prediction, which traditionally use time units instead of locations. As a result, when searching for forecasting literature you need to add the “prediction” term, which derives a much larger pool of papers, than the ones that actually do “only” forecasting. In this paper, we define the term “Spatial Crime Forecasting” as an inference approach about crime both in time and in space. In the box below, we add definition boundaries by describing variations of forecasting approaches that we consider in our study.

We are driven by the need to harmonize existing concepts and methodologies within and between criminology, sociology, geography, and computer science communities. The goal of this paper is to conduct a systematic literature review in spatial crime predictive analytics, with a focus on crime forecasting, to understand and evaluate the state of the art concerning concepts and methods given the unprecedented pace of published empirical studies. Below, we list the research questions of this study.

What are the types of forecasted information for which space plays a significant role? (“ Overview of selected publications on spatial crime forecasting ” section).

What are the commonly used forecasting methods? (“ Spatial crime forecasting methods ” section).

Which are the technical similarities and differences between spatial crime forecasting models? (“ Spatial crime forecasting methods ” section).

How is predictive performance being measured in spatial crime forecasting? (“ Considerations when analysing forecasting performance ” section).

What are the commonly used model validation strategies? (“ Considerations when analysing forecasting performance ” section).

What are the main dependencies and limitations of crime forecasting performance? (“ Considerations when analysing forecasting performance ” section).

Before presenting the results (“ Results ” section) and discuss them in the form of a SWOT analysis (“ Discussion ” section), we summarize previous literature work on crime prediction and analytics (“ Related work ” section) and then present the methodology to select the papers and ensure the study quality (“ Methods ” section). Last, in “ Conclusion ” section we conclude with the main findings of each research question. With our work, we aim to shed light on future research directions and indicate pitfalls to consider when performing spatial crime forecasting.

Related work

The papers identified as review or related-work studies (a total of 13) date back to 2003 and are connected to the keyword strategy that we used (find further details in “ Study selection ” section). In addition, to review papers (a total of 9), we also include two editorials, one book chapter, and one research paper, because they contain an extensive literature review in the field of crime predictive analytics.

Five papers focus on data mining with a much broader scope than our topics of interest, i.e., prediction, forecasting, or spatial analysis. The oldest one proposes a framework for crime data mining (Chen et al. 2004 ). It groups mining techniques into eight categories, including (a) the entity extraction (usage example: to identify persons), (b) clustering (usage example: to distinguish among groups belonging to different gangs), (c) association rule mining (usage example: to detect network attacks), (d) sequential pattern mining (usage example: same as before), (e) deviation detection (usage example: to identify fraud), (f) classification (usage example: to identify e-mail spamming), (g) string comparator (usage example: to detect deceptive information), and (h) social network analysis (usage example: to construct the criminal’s role in a network). Association rule, clustering, and classification are the ones that have been discussed in other crime data mining reviews, such as for the identification of criminals (i.e., profiling) (Chauhan and Sehgal 2017 ), applications to solve crimes (Thongsatapornwatana 2016 ), and applications of criminal career analysis, investigative profiling, and pattern analysis (with respect to criminal behaviour) (Thongtae and Srisuk 2008 ). Furthermore, Hassani et al. ( 2016 ) conducted a recent in-depth review that looked at over 100 applications of crime data mining. Their taxonomy of applications identifies five types that include those previously described by Chen et al. ( 2004 ) with the exemption of sequential pattern mining, deviation detection, and string comparator. Regarding specific algorithms, the emphasis is put on three types, namely decision trees, neural networks, and support vector machines. Chen et al. ( 2004 ) work covers a broad spectrum of crime analysis and investigation and as such, it identifies a couple of studies related to hotspot detection and forecasting under the mining categories of clustering and classification. These technical review studies gave us examples of the data items that we need to extract and analyse, such as the techniques that are used and the tasks that are performed (Thongsatapornwatana 2016 ) as well as the study purpose and region (Hassani et al. 2016 ).

The oldest, yet still relevant paper to our work is an editorial to six crime forecasting studies (Gorr and Harries 2003 ). The authors refer to crime forecasting as a new application domain, which includes the use of geographical information systems (GIS), performs long- and short-term prediction with univariate and multivariate methods, and fixed boundary versus ad hoc areal units for space and time-series data. More than 15 years later, this application domain is not new but it still involves the same characteristics as described above. Another editorial by Kennedy and Dugato ( 2018 ) introduces a special issue on spatial crime forecasting using the Risk Terrain Modelling (RTM) approach. The focus of most papers is to analyse factors that lead to accurate forecasts because the foundation of the RTM approach is based on the Theory of Risky Places by Kennedy and Caplan ( 2012 ). This theory starts with the proposition that places vary in terms of risk due to the spatial influence of criminogenic factors. Last, a recent review study summarizes past crime forecasting studies of four methods, namely support vector machines, artificial neural networks, fuzzy theory, and multivariate time series (Shamsuddin et al. 2017 ). The authors suggest that researchers propose hybrid methods to produce better results. In our study we group and discuss a much wider number of methods (a list of 66 in Additional file 1 C) and we also identified hybrid approaches (e.g., ensemble methods) one of which dates back to 2011.

In addition, we identified two papers that describe spatial methods for spatial crime prediction per se. The paper by Bernasco and Elffers ( 2010 ) discusses statistical and spatial methods to analyse crime. They interestingly distinguish two types of spatial outcomes for modelling, including spatial distribution and movement. When it comes to spatial distribution, which is relevant to the scope of our paper, the authors describe the following spatial methods, including spatial regression models, spatial filtering, geographically weighted regression, and multilevel regression with spatial dependence. The paper by Chainey et al. ( 2008 ) focuses on hotspot mapping as a basic approach to crime prediction. The techniques they describe and empirically examine are spatial ellipses, thematic mapping of geographic areas, grid thematic mapping, and Kernel Density Estimation (KDE). Among these, KDE yielded the highest prediction accuracy index (PAI) score. Surprisingly, most of the spatial methods (with the exemption of KDE and RTM) have not been used by authors of our selected papers (see methods discussed in “ Spatial crime forecasting methods ” section).

Regarding predictive policing, a recent review explains its definition, how it works, how to evaluate its effectiveness, and it also provides an overview of existing (mostly commercial) applications (Hardyns and Rummens 2018 ). One of the innovative aspects of this review is the section on the evaluation of predictive policing using three criteria, namely the correctness of the prediction, the effect of predictive policing implementations to actual crime rates, and the costs relative to the methods being replaced. The authors of this paper support the definition of predictive policing that originates from Ratcliffe ( 2015 , p. 4), which reads: “ the use of historical data to create a spatiotemporal forecast of areas of criminality or crime hot spots that will be the basis for police resource allocation decisions with the expectation that having officers at the proposed place and time will deter or detect criminal activity ”. In general, spatial crime forecasting has a broader scope and is not synonymous to predictive policing. In addition, the papers that we examine do not aim in assisting policing decisions (although this can be an indirect consequence) but they have an academic and explanatory focus. However, the effectiveness of the predictive analysis- first criterion, as framed by Hardyns and Rummens ( 2018 ), is highly connected to our scope and thus is further analysed, from a technical perspective, in “ Considerations when analysing forecasting performance ” section.

A second predictive policing systematic review by Seele ( 2017 ) examines the potential of big data to promote sustainability and reduce harm and also discusses ethical and legal aspects linked to predictive algorithms. Similarly, Ozkan ( 2018 ) also reviews big data for crime research. This paper provides a critical discussion on the benefits and limitations of data-driven research and draws attention to the imminent threat of eliminating conventional hypothesis testing, which has traditionally been an integral scientific tool for social scientists and criminologists.

Except for Seele ( 2017 ) no other related-work study follows a systematic procedure regarding the methods to identify and select relevant research, and thereafter to collect and analyse data from them. Also, our work focuses only on spatial crime forecasting, which is narrower than crime data mining and broader than predictive policing as discussed above. Last, we aim to contribute with scientific reference for technical issues in future studies. To achieve this, we follow a review protocol (“ Methods ” section), to answer six research questions (mentioned in “ Background ”) that have not been systematically addressed by previous studies.

Study selection

This study follows the reporting guidance “PRISMA” (Preferred Reporting Items for Systematic reviews and Meta-Analyses) (Liberati et al. 2009 ). PRISMA suggests a checklist of 27 items regarding the sections of a systematic literature review and their content, as well as a four-phase flow diagram for the selection of papers. We adopted and modified the PRISMA guidance according to the needs of our study. Our flow diagram contains three phases for the selection of papers. The first phase is “identification” and involves the selection of information sources and a search strategy that yields a set of possible papers. The second phase is “screening” the selected papers from the first phase, and removing the ones that are not relevant to the research scope. The third phase is “eligibility”, in which we applied a more thorough reading of papers and selected the ones that are relevant to our research questions. The count of papers in each phase and their subsequent steps are illustrated in Fig.  1 .

The three phases of the study selection process: identification, screening, and eligibility

The number of papers selected in the Identification phase is based on eleven keywords related to crime prediction (i.e., predict crime, crime predictive, predictive policing, predicting crimes, crime prediction, crime forecasting, crime data mining, crime mining, crime estimation, crime machine learning, crime big data). In addition, we added seven more spatially explicit terms (i.e., crime hotspot, spatial crime prediction, crime risk terrain modelling, spatial crime analysis, spatio-temporal modelling crime, spatiotemporal modelling crime, near-repeat crime). In a subsequent analysis, we have visualized the word frequency of the titles of the selected papers as evidence of the relevance of the keywords used. This can be found in Additional file 1 B: A word cloud of high - frequency words extracted from the titles of the selected papers .

Next, we selected information sources to perform literature searches. Although there is a remarkable number of search engines and academic databases, we focus on scholarly and comprehensive research databases including fields where spatial crime prediction is a representative topic. We considered the following databases, including Web of Science by Clarivate Analytics (WoS), Institute of Electrical and Electronics Engineers (IEEE) Xplore, ScienceDirect by Elsevier (SD), and Association for Computing Machinery (ACM) Digital Library. We consider that an optimal search process should include multiple academic search databases, with searches being carried out at the best level of detail possible. In addition, as also discussed by Bramer et al. ( 2017 ) in an exploratory study for database combinations, if the research question is more interdisciplinary, a broader science database such as Web of Science is likely to add value. With regards to Google Scholar (GS) there are divergent opinions between researchers if GS brings relevant information for an interdisciplinary review or not. Holone ( 2016 ) discusses that some engine searches, specifically GS, have a tendency to selectively expose information by using algorithms that personalize information for the users, calling this the filter bubble effect. Haddaway et al. ( 2015 ) found that when searched for specific papers, the majority of the literature identified using Web of Science was also found using GS. However, their findings showed moderate to poor overlap in results when similar search strings were used in Web of Science and GS (10–67%), and that GS missed some important literature in five of six case studies.

In each database, we used keywords on singular and plural word versions (e.g., crime hotspot/s). For WoS, we used the advanced search option, by looking for papers written in English and matching our keywords with the topic or title. For IEEE, we searched for our keywords in the metadata or papers’ titles. In SD and ACM, we used the advanced search option with Boolean functions that searched our keywords in the title, abstract, or paper’s keywords. The identified papers were integrated directly into the free reference manager Mendeley. Last, we removed duplicates within each database, which resulted in 786 papers for the second phase, the Screening phase. The last search in the Identification phase was run on 5 February 2019.

Whereas, the use of statistical and geostatistical analyses for crime forecasting has been considered for quite some time, during the last two decades there has been an increasing interest in developing tools that use large data sets to make crime predictions (Perry 2013 ). Thus, predictive analytics have been included in law enforcement practices (Brayne 2017 ). This is the main reason that during the Screening phase, we first excluded papers published before 2000. Second, we removed duplicates across the four selected databases (WoS, IEEE, SD, and ACM). Third, we screened all papers to identify the “non-relevant” ones. This decision was made by defining “relevant” papers to contain the following three elements. The first element is that a paper addresses crime events with explicit geographic boundaries. Common examples of excluded papers are the ones dealing with the fear of crime, offenders’ characteristics, offender, victims’ characteristics, geographical profiling, journey to crime, and cyber or financial crime. The second element for a paper to be “relevant” is that it employs a forecasting algorithm and is not limited to exploratory or clustering analysis. The third element is that there is some form of spatial prediction. This means that there are predefined spatial units of analysis, such as inferencing for each census block of the study area. For the relevance elements, our strategy was the following: (a) read title and screen figures and/or maps, (b) if unsure about relevance, read abstract, (c) if still unsure about relevance, search for relevant words (e.g., geo*, location, spatial) within the document. The last step of the Screening phase was to remove relevant papers that authors did not have access to, due to subscription restrictions. The Screening phase resulted in 193 relevant papers to be considered for the third and final phase.

During this final phase, the Eligibility phase, we read the abstract and main body of all 193 papers (e.g., study area, data, methods, and results). The focus was to extract data items that compose the paper’s eligibility criteria. These are grouped into three categories, namely: (a) publication type which is the first data item, (b) relevance: consists of data items relevance and purpose , and (c) study characteristics: consists of data items study area , sampling period , empirical data , evaluation metrics . Next, we discuss each category and the data items it entails.

The first data item is the publication type. Literature reviews sometimes exclude conference papers because their quality is not evaluated like International Scientific Indexing (ISI) papers. However, for some disciplines, such as computer science, many conferences are considered as highly reputable publication outlets. In the Screening phase, we found a large number of papers from computer or information science scholars, hence at this stage we decided not to exclude conference papers (n = 65), but also non-ISI papers (n = 19). In total, we excluded nine papers that are book chapters or belong to other categories (e.g., editorial).

The next two “relevance” criteria (i.e., relevance and purpose) address the fit of the papers’ content to our research scope. Paper relevance was checked again during this phase. For example, some papers that appeared to be relevant in the Screening phase (i.e., a paper is about crime events, space, and forecasting), were actually found not to be relevant after reading the core part of the paper. For example, prediction was mentioned in the abstract, but what the authors implied was that prediction is a future research perspective of the analysis that was actually done in the paper. Also, we added the data item “purpose” to separate methods that model and explore relationships between the dependent and independent variables (e.g., crime attractors to burglaries) from the ones that perform a spatial forecast. The number of papers that were excluded due to these criteria amounted to 81.

Last, there are four more “study characteristics” criteria relevant to the quality and homogeneity of the selected papers. First, the study area should be equal to or greater than a city. Cities are less prone to edge effects compared to smaller administrative units within a city that share boundaries with other units (e.g., districts). In addition, the smaller the study area the more likely it is that conclusions are tailored to the study characteristics and are not scalable. Second, the timeframe of the crime sample should be equal or greater than a year to ensure that seasonality patterns were captured. These two items also increase the homogeneity of the selected studies. Yet, there are significant differences among studies that are discussed further in Results section. The last two criteria are the restriction to analysing empirical data (e.g., proof of concepts or purely methodological papers were excluded) and to use measures that evaluate the models’ performance (e.g., mean square error). The last two criteria ensure that we only analyse studies that are useful to address our research questions. The number of papers that were excluded due to the publication type, the relevance, or the study characteristics were 71. Furthermore, Fig.  1 shows the number of excluded papers for each data item (e.g., 17 papers were excluded due to insufficient size of the study area). Finally, the entire selection processes yielded 32 papers.

Study quality

Two of the four authors of this research performed the selection of the papers to be analysed. Prior to each phase, these two authors discussed and designed the process, tested samples, and divided the workload. Then, results were merged, analysed, and discussed until both authors reached a consensus for the next phase. The same two authors crosschecked several of the results to ensure methodological consistency among them. The reading of the papers during the final phase (i.e., eligibility) was performed two times, by alternating the papers’ samples among the two authors, to ensure all eligible papers were included. In addition, in case some information on the paper’s content was unclear to the two authors, they contacted via email the corresponding authors for clarifications.

Regarding the results subsections that constitute four study stages (“ Study characteristics ”, “ Overview of selected publications on spatial crime forecasting ”, “ Spatial crime forecasting methods ”, and “ Considerations when analysing forecasting performance ” sections), one or two authors performed each and all authors contributed to extracting information and reviewing them. To extract information that is structured as data items we followed a procedure of three steps that was repeated at each stage. First, the papers were read by the authors to extract manually the data items and their values (1—extract). Data items and their values were then discussed and double-checked by the authors (2—discussion/consensus). In case information was still unclear, we contacted via email the corresponding authors for clarifications (3—consultation). This information was structured as a matrix where rows represent the papers and columns are several items of processed information (e.g., a data item is the year of publication). Table  1 shows the data items at the stage at which they were exploited. The attributes (values) of the items are discussed in “ Results ” section.

The risk of bias in individual studies was assessed via the scale of the study. Spatial and temporal constraints were set (already defined in the eligibility phase) to ensure that we analyse medium to large scale studies and that findings are not tied to specific locality or seasonality characteristics. Furthermore, we did not identify duplicate publications (i.e., two or more papers with the same samples and experiments) and did not identify study peculiarities, such as special and uncommon features or research topics.

Last, the risk of bias across studies was assessed via an online survey. We contacted the authors of the publications (in some cases we could not identify contact details) and ask them to respond to a short survey regarding the results of their paper. The introductory email defined the bias across studies as “ Bias across studies may result from non - publication of full studies (publication bias) and selective publication of results (i.e., selective reporting within studies) and is an important risk of bias to a systematic review and meta - analysis” . Then, we explained the content of the survey that is to identify, if there are non-reported results that are considerably different from the ones in their papers. This information was assessed via two questions (for further details we added the questionnaire as a Additional file 1 of this paper). Out of the 32 papers, we received responses for 11 papers ( n  =  12, with two authors responding to the same paper). One factor that explains the low response rate is that many authors have changed positions (papers date back to 2001) and for some we could not identify their new contact details, while for others we received several non-delivery email responses.

Regarding the responses’ results, seven authors responded that they never conducted a similar study to the one for which they were contacted for and five responded that they have conducted a similar study to the one for which they were contacted. A similar study was defined as a study in which: (a) the study design, selection of independent variables/predictors, selection of method(s), and parametrization of a method(s) are the same, and (b) data can be different. From those who performed a similar study four responded that their results were not different and one responded that their results were considerably different. However, in a follow-up explanatory answer, this author responded that changing the parametrization yielded different results regarding the performance ranking of three algorithms and that the data and the study area were the same. Based on this small-scale survey there is no indication that there is a risk of bias across studies. However, further investigation of this matter is needed.

Study characteristics

In this section, we discuss generic characteristics of the selected papers. To start with, the type of publication is slightly higher for ISI journal articles (n = 18) than for conference papers (n = 14). The 32 papers were published in a variety of journals and conferences and no preference was observed for a particular publication outlet. In specific, four journals and one conference published two or three of the selected papers each (Table  2 ) and all other papers were published in different journals and conferences. On the other hand, there is little variation regarding the country of the study area. The majority of studies were conducted in the US, which is probably a somewhat biased statistic, considering the large US population size, as well as the used language (e.g., English) of the study selection process. Similarly, institutions that have published more than one paper on spatial crime forecasting are based in the US with the exception of the Federal University of Rio Grande do Norte, Brazil, that has recent publications in this field.

We also collected the discipline associated with each paper. To do so we used the affiliation of the first author and extracted the discipline down to the department level, if this was possible. Then we used as a benchmark reference the 236 categories/disciplines used in Journal Citation Reports (JCR) Footnote 1 by the Web of Science Group. Each affiliation of authors was then matched to one of the categories. In Table  2 , we see the disciplines that appeared more than one time (i.e., computer science, criminology, public administration, and geosciences). Although we collected a variety of disciplines these are the ones that we encountered more than once and account for the majority of the papers ( n  = 22). Thus scholars of these disciplines seem to have a greater interest in spatial crime forecasting.

Figure  2 shows the number of eligible and selected articles by year during the study selection period. We included the eligible in addition to the selected papers for two reasons. First, many of the eligible papers looked into spatial crime forecasting but did not meet the criteria defined for this study. Second, other papers may not be relevant to forecasting, but are relevant to the broader topics of prediction or modelling. The graph in Fig.  2 depicts a rapidly increasing trend over the last couple of years. For the eligible papers, the number of articles increased substantially since 2013, whereas for the selected papers, a similar trend is evident in the last 2 years.

A yearly count of eligible and selected papers from 2001 to 2018

Overview of selected publications on spatial crime forecasting

In Table  3 we enlist each selected paper along with information related to space (i.e., study area and spatial scale), time (i.e., sampling period and period in months), crime data (i.e., crime type and crime sample size), and prediction (i.e., predicted information, task, spatial unit, and temporal unit). In this section, we consider these 10 data items as initial and basic information when reporting a spatial crime forecasting study. A reader who may want to replicate or modify the methodological approach presented in the follow-up research will require the same 10 data items to assess whether such approach is adequate to the author’s follow-up study and research objectives. More importantly, when any of these data items are missing an assessment of the generalizability (or bias) of the conclusions and interpretation of results is limited. Unfortunately, the majority of the 32 selected papers (n = 21) had at least one item with undefined or unclear information for five out of the 10 data items (Fig.  3 ). From these, 52% (n = 11) were conference papers and 48% (n = 10) were ISI articles. On the other hand, 73% (n = 8) of the papers with no undefined or no unclear information were ISI papers and 27% (n = 3) were conference papers.

Percentages of all publications (n = 32) for describing basic information when reporting a spatial crime forecasting study. Blue: the item was properly defined; orange: the item was poorly defined or undefined

Most of the studies were conducted at the city level. In two studies, the forecasting area covered a county, which is the US administrative unit that usually expands across a city’s boundary. In one paper, predictions covered an entire country (US). New York City (US) was examined in four studies, Pittsburgh (US) was examined in three studies, and Portland (US), Natal (Brazil), and Chicago (US), were examined in two studies. All other publications were based on individual study areas.

The oldest crime data that were used in the 32 papers are from the year 1960 and the most recent crime data are from 2018. The sampling period ranges from 1 year up to 50 years. There is one paper with an undefined sampling period. However, from the article it can be inferred that the length of the sampling period is at least 1 year. Regarding the sample size of the crime data, it ranges from about 1400 incidents up to 6.6 million, which is relevant to the number of crime types as well as to the length of the sampling period. As for the number of crime types, there are four studies that aggregated and analysed all crime types together, twelve studies that focused on a particular crime type, fourteen studies that looked from two to up to 34 different crime types, and three studies with undefined information on the crime type analysed. Residential burglary was the crime type that was most often examined in studies that looked into only one crime type.

The last four data items in Table  3 describe details of the forecasted information, which we refer to as “inference”. The temporal unit is the time granularity of the inference and ranges from a fine granularity of 3 h up to 1 year. The most frequent temporal unit across all papers is 1 month (used in 12 papers). In addition, day and week have been used in eight studies and years in seven studies. Other less frequent temporal units are 3 h, daytime for 1 month, night-time for 1 month, 2  weeks, 2 months, 3 months, and 6 months. Similarly, the spatial unit is the space granularity of the inference and ranges from a small area of 75 m × 75 m grid cell size to a large area, such as police precincts or even countries. The spatial unit is difficult to analyse and to compare for two reasons. First, spatial units do not have a standard format like time and are two-dimensional. Thus, they can vary in size and shape. Second, for about one-third of the papers this information was poorly reported (Fig.  3 ). In the case of administrative units (e.g., census blocks or districts), the shape and size usually vary, but if someone is looking for further details or the units themselves, these can be in most cases retrieved by administrative authorities. However, spatial clusters may also vary in shape and size, but if details are not reported (e.g., the direction of ellipses, the range of clusters’ size, the average size of clusters) it is difficult to quantify and replicate such clusters. We also encountered cases where authors report dimensions of a grid cell size without mentioning the units of measurement. Nevertheless, the grid cell seems to be the preferable type of spatial unit and it is used in the majority of papers (n = 20).

The data items “inference” and “task” refer to the types of forecasted information and predictive task, respectively. Inference and task are defined according to the information that the authors evaluated and not to the output of a prediction algorithm. For example, an algorithm may derive crime intensity in space (i.e. the algorithm’s output), which the authors used to extract hotspots (i.e. processed output to be evaluated) from and then evaluate their results using a classification measure such as precision, accuracy, or others. Some predictive methods, such as random forest, can be used for both classification and regression tasks. It is unclear why some authors choose to apply a regression application of a method and then process, derive, and evaluate a classification output, although they could do this by directly applying a classification application of the same method. In addition, the inference “hotspots” in Table  3 includes the next four categories:

Concerning categories three and four, some authors refer to these areas as hotspots and others do not. We group all four categories together and define them as hotspots and non-hotspots, representing the output of a binary classification that separates space into areas for police patrolling that are alarming and non-alarming. We acknowledge that in the field of spatial crime analysis, hotspots are areas of high crime intensity. However, in our selected papers there does not seem to be a clear definition of the term “hotspot”.

The majority of the papers (n = 20) inferred hotspots as the outcome of a binary classification. Nine studies inferred the number of crimes or the crime rate in each spatial unit. However, three studies appear to be somehow different and unique from the rest. Huang et al. ( 2018 ) evaluated the forecasted category of crime as the output of a binary classification problem (e.g., is category A present in area X; yes or no). Ivaha et al. ( 2007 ) inferred the total number of crimes in a region, spatial clusters (or hotspots), and the share of crime within each cluster. Last, Rodríguez et al. ( 2017 ) evaluated the properties (i.e., location and size) of inferred clusters.

Spatial crime forecasting methods

The first three data items that were extracted to be analysed in this section are the proposed forecasting method, best proposed forecasting method , and the baseline forecasting method . The latter is the method used as a comparison measure of the proposed method. We analysed the frequency of the methods for each of the three forecasting types. The best proposed forecasting method is the one with the best performance throughout the conducted experiments. For example, if an experiment is evaluated separately on multiple types of crimes, we only count the method with the best performance for most cases. In case two methods perform similarly (as evidenced by statistical results and reported by the authors of the papers), both methods are considered. This was necessary because some papers proposed more than one method to be compared with a baseline method, but in the end, these papers propose the method with the best performance. In addition, this reduces biased frequency counts of proposed methods. On the other hand, we considered as a baseline the method, with which the authors wanted to compare the proposed methods. For instance, Zhuang et al. (2018) proposed three Deep Neural Networks and used an additional six machine learning algorithms as baseline methods to assess how much better the proposed methods were compared to the baseline methods. In this case, we counted the six machine learning algorithms as the baseline methods.

In Table  4 , we show “top” methods (i.e., frequently counted within the 32 selected papers) by each item. Random Forest (RF) is the most frequently used proposed method. Multilayer Perceptron (MLP) appears as a top method in all three items (i.e., proposed, best, baseline). Other best proposed methods are Kernel Density Estimation (KDE)-based and Risk Terrain Modelling (RTM). Interestingly, Support Vector Machines (SVM) have been proposed in five papers, but are not among the top four best-proposed methods. On the other hand, plenty well-known statistical models, are preferred as baseline methods, such as Autoregressive model (AR)-based, Logistic Regression, Autoregressive Integrated Moving Average model (ARIMA), and Linear Regression, as well as KDE-based and K Nearest Neighbours. In Additional file 1 C we added detailed tables that show for each paper the data items proposed method, best proposed method, and baseline method.

In the next sections, we categorize the proposed forecasting methods by type of algorithm (“ Algorithm type of proposed forecasting methods ” section) and by type of inputs they take (“ Proposed method input ” section). This task was challenging because there is no scientific consensus on a single taxonomy or categorization of analytical techniques and methods. Papamitsiou and Economides ( 2014 ) reviewed papers in educational analytics, categorizing data mining methods into classification, clustering, regression, text mining, association rule mining, social network analysis, “discovery with models”, visualization, and statistics. Other researchers would summarize all of these categories, for instance, as supervised learning, non-supervised learning, and exploratory data analysis. Vlahogianni et al. ( 2014 ) use different categorizations for reviewed papers in traffic forecasting, including aspects related to the model’s approach to treating inputs and other properties relevant to split the proposed methodologies. The right granularity of properties to define a useful categorization can be problematic and particular for each field.

Algorithm type of proposed forecasting methods

Another suitable characteristic to classify forecasting methods is the similarities between algorithms. We divide all algorithms used in the reported papers into (i) kernel-based (ii) point process, (iii) traditional machine learning, and (iv) deep learning, according to the following criteria. Kernel-based algorithms are particularly concerned with finding a curve of crime rate \(\lambda\) for each place \(g\) that fits a subset of data points within the boundaries of a given kernel (see Eq.  1 ). We observe that the main difference among kernel-based algorithms is the use of different kernel types. Hart and Zandbergen ( 2014 ) experimented with different kernel types, providing some useful conclusions. In our selected papers, six of them have used kernel-based algorithms with the most frequently used the simple two-dimensional Kernel Density Estimation (KDE) (n = 2). However, we observed that some methods are a variation from the simple KDE model, in the form of the Spatio-Temporal KDE (STKDE) used in the paper by Hu et al. ( 2018 ), the Network-Time KDE (NTKDE) proposed by Rosser et al. ( 2017 ), or the dynamic-covariance KDE (DCKDE) proposed by Malik et al. ( 2014 ). We also have considered the Exponential Smoothing method used in the paper of Gorr et al. ( 2003 ) as a kernel-based algorithm, since it works with a window function (kernel) on time series aggregation.

Point processes can be distinguished from kernel-based algorithms insofar as a background rate factor \(\mu\) that can be calculated stochastically, such as with a Poisson process, is considered. The background rate factor includes the modelling of covariates or features of the place \(g\) , such as demographic, economical, geographical, etc. variables (see Eq.  2 ). From the explanation made by Mohler ( 2014 ), we suppose that the introduction of the background rate makes the point process more suitable for multivariate modelling when compared to kernel-based algorithms. In the reviewed papers, algorithms can be distinguished among each other based on their mathematical formulations of \(\kappa\) and \(\mu\) , but also on their internal parameter selection, mostly based on likelihood maximization. Only three papers proposed such an algorithm, including the Marked Point Process from Mohler ( 2014 ), the maximum likelihood efficient importance sampling (ML-EIS) from Liesenfeld et al. ( 2017 ), and the Hawkes Process from Mohler et al. ( 2018 ).

In the case of machine learning algorithms, their formulation is often associated with finding a function \(f\) that maps feature vectors X to a given output Y. These algorithms are distinguished from each other in the way this function is estimated, some being more accurate and complex than others. We include in this category all algorithms that are explicitly associated with regression or classification. They differ from algorithms of previous categories, because \(f\) is constructed only after a training process. This training step aims to find a model that minimizes the estimation error between the predicted output and the original output. The majority of the reported papers (n = 20) was included in this class of algorithms. The most proposed traditional machine learning algorithms were RF and MLP (tied at n = 6), followed by SVM together with Logistic Regression (n = 4), and Negative Binomial Regression used in RTM studies together with (n = 3).

Although deep learning algorithms have the same formulation as traditional machine learning algorithms, they present a much more complex internal structure that affects their use. The deep layer structure makes the computational budget mainly needed during training. Additionally, the need for samples is also greater, than for the other approaches. Among the reported papers, the three that have used this type of algorithm argue that it has the best overall performance. This includes the Deep Neural Networks (DNN) fitted by Lin et al. ( 2018 ), the DeepCrime framework from Huang et al. ( 2018 ), and the Long Short-Term Memory (LSTM) architecture proposed by Zhuang et al. ( 2017 ). The paper by Huang et al. ( 2018 ) even presents a neural architecture dedicated to a feature-independent approach, with a recurrent layer dedicated to encoding the temporal dependencies directly from the criminal occurrences. Still, none of these papers has discussed computational time performance against other algorithms, nor sample sizes sufficient to obtain accurate models. At the time of this writing, we argue that there is no clear guidance on when one should conduct a deep neural networks approach, although in recent years evidence of its effectiveness has begun to emerge.

Proposed method input

Another split factor is the inputs of the forecasting methods, i.e. the independent variables. There are some forecasting methods that accept as input the latitude, longitude, and timestamp of criminal events (raw data), while others need to apply explicit aggregations or transformations before feeding their models. In this paper, we refer as feature engineering the process of crafting, scaling and selecting covariates or features to better explain a prediction variable which often requires considerable domain-specific knowledge. An example is the aggregation of criminal events into spatiotemporal series, which can be decomposed into autoregressive lags and used as features. This feature engineering can also be applied to ancillary data methodologies not directly related to crime. For instance, Lin et al. ( 2018 ) count landmarks on the grid by counting the number of items in each cell (spatial aggregation) and craft a new feature for each landmark type, while Huang et al. ( 2018 ) define a part of their algorithm being a region embedding layer for only processing the raw location of the city’s landmarks. We believe that the split factor by method inputs may be useful information for a potential researcher who wishes to perform spatial forecasting and consults this section of our paper. Data processing requires domain knowledge, and it is an expensive (timewise) task, especially when dealing with large spatiotemporal datasets. Thus, avoiding the feature-engineering process may be preferable by some researchers. On the other hand, one may prefer to use data to derive their variables with particular patterns.

We call methods that have an internal approach to aggregating crime events into spatiotemporal variables “feature-engineering independent” and “feature-engineering dependent”. In other words, these methods explicitly need aggregations to derive spatiotemporal variables from the raw data independently of the forecasting algorithm. The majority (n = 24) of reported papers have an explicit way to transform their crime events, as well as ancillary data, into features to feed their algorithms (i.e., feature-engineering dependent). Although we have found many different forms of data aggregation into features, both spatially and temporally, the procedure of assigning features is often insufficiently reported, making it difficult to reproduce the proposed methodology. Still, well-defined workflows or frameworks followed by feature-engineering dependent methods were detailed in Malik et al. ( 2014 ) and Araújo et al. ( 2018 ). They synthesized their forecasting methods in (1) aggregate raw data spatially, following a crime mapping methodology (e.g., counting events inside grid cells), (2) generate time series and their features, (3) fit a forecasting model using an algorithm, and (4) visualize the results. In feature-engineering dependent methods the aggregation and time series generation is done separately as processing steps before fitting a model, whereas this is not needed for the feature-engineering independent methods.

Considerations when analysing forecasting performance

In this section, we look at measures of forecasting performance (“ Overview of evaluation metrics ” section) and discuss which are used for each forecasting task, including for classification and regression (“ Metrics by forecasting task ” section). Then, we explore validation strategies by types of algorithms (“ Algorithms and validation strategies ” section). Finally, we summarize and discuss the main dependencies and limitations of the above subsections (“ Dependencies and limitations ” section).

Overview of evaluation metrics

As mentioned in “ Spatial crime forecasting methods ” section, selected papers include forecasting baseline models, novel models, or ensemble models proposed by respective authors. Evaluation metrics of such models are in general, well-known in criminology, GIScience, mathematics, or statistics. However, it is important to mention that few authors highlight the necessity of combining or using diverse evaluation metrics.

We cannot make a comparison of all evaluation results across the 32 papers due to various reasons, such as different spatial and temporal units, study areas, or forecasting methods applied. Yet, we can discuss certain similarities between them. Choosing an evaluation metric is highly dependent on the main prediction outcome, such as counts (e.g., for a Poisson distribution), normalized values or rates (e.g., for a continuous Gaussian distribution), or binary classification (crime is absent or present). The most frequent evaluation metrics used in the selected papers are the Prediction Accuracy (PA, n = 9), followed by the Prediction Accuracy Index (PAI, n = 7), the F1-Score (n = 7), Precision and Recall (n = 5), the Mean Squared Error (MSE, n = 4), the Root Mean Squared Error (RMSE, n = 3), R-squared (n = 3), the Recapture Rate Index (RRI, n = 3), the Hit Rate (n = 2), the Area Under the Curve (AUC, n = 2), and the Mean Absolute Forecast Error (MAFE, n = 2). Some additional metrics are used only once, namely the Spatio-Temporal Mean Root Square Estimate (STMRSE), the average RMSE (RMSE), the Regression Accuracy Score (RAS), the Regression Precision Score (RPS), the Ljung-Box test, the Mean Absolute Error (MAE), the Mean Absolute Percentage Error (MAPE), macro-F1, micro-F1, the Mean (Squared) Forecast Error (M(S)FE), the Pearson Correlation Coefficient, and the Nash coefficient. Generally, metrics derived from the confusion matrix, namely accuracy, precision, recall, and F1-Score, are used together to evaluate binary classifications.

We analysed the top three evaluation metrics (PA, PAI, F1-Score) in relation to their distribution among the data items of discipline, proposed forecasting algorithm type, forecasting inference, forecasting task, spatial unit, and temporal unit. Interestingly, we found that computer scientists exclusively use the PA, while criminologists prefer to apply the PAI. In addition, while the PA and the F1-Score have been preferably tested for short-term predictions (i.e., less or equal to 3 months), the PAI has been used for both short and long-term predictions. No other obvious pattern was detected among the other information elements regarding the usage and preference of these evaluation metrics.

Metrics by forecasting task

The most common forecasting task is binary classification (n = 21) for crime hotspots (n = 20) and the category of crime (n = 1). While the classification task is frequently discussed at the beginning of experiments, some articles consider in the performance evaluation a different item than in the output of the algorithm, thus transforming regression products into binary values. The most prominent examples include RTM models (Drawve et al. 2016 ; Dugato et al. 2018 ; Gimenez-Santana et al. 2018 ), where the output of the algorithm is a risk score. This score is later reclassified into a binary outcome (a positive or negative risk score) for the purpose of the evaluation. In addition, Rummens et al. ( 2017 ) propose a combined ensemble model consisting of LR and MLP that infers risk values, similar to RTM, where authors consider as crime hotspot, values with a risk higher than 20%.

The regression task (n = 11) is largely used for predicting the number of crimes (n = 8) and the performance is measured by various error measurements, such as MSE (n = 4) or RMSE (n = 3). Araujo et al. ( 2017 ) propose two new evaluation metrics, namely the Regression Accuracy Score (RAS), representing the percentage of success in predicting a sample, and the Regression Precision Score (RPS), which defines the RAS’s precision. The RPS measures the MSE of success samples normalized by the variance of the training sample (Araujo et al. ( 2017 )). Rodríguez et al. ( 2017 ) introduce the Nash–Sutcliffe Efficiency (NSE), which they derive from hydrological models forecasting, as a normalized statistic determining the relative magnitude of the residual variance compared to the measured data variance.

However, the number of crimes is not the only inference considered in regression models. For example, Ivaha et al. ( 2007 ) predict the percentage of crime in clusters, using spatial ellipses as spatial units, Rodríguez et al. ( 2017 ) investigate properties of clusters, while Shoesmith ( 2013 ) infers crime rates from historical crime data.

In addition to the above-mentioned evaluation metrics, three articles discuss surveillance plots for prediction evaluation. Mohler ( 2014 ) uses a surveillance plot metric showing the fraction of crime predicted over a time period versus the number of grid cells with real crimes for each day (Fig.  4 a). The same author mentions that this metric is similar to the receiver operating characteristic curve, or ROC curve, applied by Gorr ( 2009 ), but differs because it is not using an associated false positive rate on the x-axis. Similarly, Hu et al. ( 2018 ) apply the PAI curve, also a surveillance plot showing the area percentage on the x-axis, and the PAI or the hit rate value on the y-axis (Fig.  4 b, c). Similarly, Rosser et al. ( 2017 ) use hit rate surveillance plots, representing the mean hit rate against the coverage for the network and grid-based prediction approaches (Fig.  4 c). These plots are highly useful to visualize metrics’ values over the surveyed territory.

Comparable surveillance plots for evaluation metrics visualization in space (using dummy data). a ROC-like accuracy curve, b PAI curve, and c Hit rate curve

Algorithms and validation strategies

As mentioned in “ Spatial crime forecasting methods ” section, in many of the papers, the proposed forecasting method does not include a novel algorithm, but mostly applies new variables that have not previously been used before. When reminding us of the four types of algorithms, namely (i) kernel-based, (ii) point process, (iii) traditional machine learning, and (iv) deep learning, we note a diversity between the proposed forecasting and the baseline methods. In addition, validation strategies are diverse, as well. Half of the studies (n = 16) consider splitting the data into training and testing subsets. Most of these studies include 70% training (current) for 30% testing (future) sets. Johansson et al. ( 2015 ) use a combined approach, including rolling horizon, which is producing ten times the size of a sample for the KDE algorithm, containing 70% of the original crime dataset (keeping the 70/30 ratio). The final result is calculated as the mean of the ten measurements. Figure  5 gives a good overview of all algorithms and their validation strategies. This decision tree visualization includes five central data items, namely prediction task, proposed input forecasting method, proposed forecasting algorithm type, validation strategy, and evaluation metrics. Classification m refers to those evaluation metrics that are particularly used for classification tasks (e.g., PA, F1-score). Regression m is a composition of error metrics for regression analysis (e.g., MSE, RMSE, MAE), while Criminology m includes crime analysis metrics (e.g., PAI, RRI).

Overview of forecasting methods (see “ Spatial crime forecasting methods ” section) and their performance evaluation (see “ Considerations when analysing forecasting performance ” section) linked to the 32 selected papers. The papers’ references linked to their number are shown in Table  3 . The letter m denotes an evaluation metric. The letter “U” denotes an undefined item

Kernel-based algorithms are preferably used to predicting hotspots (n = 5) and the number of crimes (n = 1). Interestingly, Malik et al. ( 2014 ) bring into discussion the fact that regions with similar demographics tend to illustrate similar trends for certain crime types. This observation is included in their prediction model “Dynamic Covariance Kernel Density Estimation method (DSKDE)” and compared with the “Seasonal Trend decomposition based on Loess (STL)” baseline model. Hart and Zandbergen ( 2014 ) and Johansson et al. ( 2015 ) use a kernel-based KDE approach without comparing it with a baseline method, both considering the PAI as one of the evaluation metrics. Only two of the kernel-based studies consider ancillary data (Gorr et al. 2003 ; Rosser et al. 2017 ), yet both use different validation strategies (rolling-horizon and train-test split, respectively) and evaluation metrics (MAE, MSE, MAPE in the first publication and Hit Rate in the second publication). Thus, it is worth noting that, while using the same base algorithm, such as KDE, other components of the prediction process may be different.

Two out of three point process algorithms do not explain the validation strategy followed in the studies (Liesenfeld et al. 2017 ; Mohler 2014 ). Mohler ( 2014 ) shows an interesting point process approach using only historical crime data, capturing both short-term and long-term patterns of crime risk. This article includes the surveillance plot evaluation (see “ Metrics by forecasting task ” subsection), comparing chronic and dynamic hotspot components for homicides and all crime types.

The third category of forecasting algorithms, the traditional ML, is split up almost equally between classification and regression tasks. Only three articles discussing traditional ML algorithms do not mention information about the baseline comparison (Araújo et al. 2018 ; Rodríguez et al. 2017 ; Rummens et al. 2017 ). The majority of ML algorithms (n = 11) use the training–testing split validation strategy applied to the classification task. Interestingly, one of the articles (Yu et al. 2011 ) discusses a different validation approach, the “Leave-One-Month-Out” (LOMO), where instead of running the classification only once on the training and testing data sets, it is run on S − 1 sets (S = number of sets/months).

An increasing body of forecasting techniques are based on DL, however, for this review, we include only three articles, with all of them for short-term prediction and coming from the computer science discipline (Huang et al. 2018 ; Lin, Yen, and Yu 2018 ; Zhuang et al. 2017 ). Two of the three articles consider geographic ancillary variables and apply the rolling-horizon validation strategy, while the third article deals only with crime lags following a 10-fold cross-validation approach. All three articles consider a binary classification evaluated by metrics such as the PA and the F1-score. Zhuang et al. ( 2017 ) propose a spatio-temporal neural network (STNN) for forecasting crime occurrences, while embedding spatial information. They then compare the STNN with three state-of-the-art methods, including the Recurrent Neural Network (RNN), the Long Short-Term Memory (LSTM), and the Gated Recurrent Unit (GRU). Since the model is designed for all types of crime data, each crime type can lead to different performances of the STNN due to their variability in time and space. Presumably, challenges will appear for crime types with low data volumes, because neural networks require a sufficient amount of data for training.

Dependencies and limitations

Although most papers use standard evaluation metrics, such as PA for a binary outcome, they usually do not include complementary metrics, in order to ensure that every aspect of the prediction performance is covered. Often, the PA is used by itself to measure model performance (Araújo et al. 2018 ; Malik et al. 2014 ; Mu et al. 2011 ). Complementary metrics are needed, because whilst one method may have a higher evaluation score than others, they may provide additional information. For example, while showing a high PAI, the Prediction Efficiency Index (PEI) value (Hunt 2016 ) may be reduced. PEI is another evaluation metric that is calculated by the ratio of the PAI to the maximum possible PAI a model can achieve. The difference between the PAI and the PEI can be explained by both metrics having different dependencies on the cell size.

Complementary metrics also overcome limitations of some evaluation measures. For example, the PA is the sum of true positives and true negatives divided by the total number of instances, which represents the percentage that is classified correctly. However, this information may not be enough to judge the performance of a model, because it omits information about precision. The Hit rate and the PAI are obtained through a division, thus, when the denominator is small, both metrics are high. Consequently, when crime occurrences are low, results are heavily affected.

Furthermore, traditional metrics are global in nature, but in spatial prediction or forecasting, we are also interested in the spatial distribution of the prediction. There may be local areas of good and bad prediction performance, resulting in an average global value. A complementary metric for a regression outcome could be to calculate the Moran’s I of the prediction error and explore the variation of the predictive performance throughout the study area. Ideally, the prediction error should follow a random spatial distribution. Overall, we find a low to no interest in developing (local) spatial, temporal, or spatiotemporal evaluation metrics.

The relevance of evaluation metrics may be biased for various reasons. One example can be the class imbalance. A model can have high accuracy while predicting the locations without crime very well. In contrast, locations with crimes are not well forecasted. Some authors try to ameliorate the negative–positive ratio between crime and no crime cells, by adjusting the weight of hotspots and cold spots (Yu et al. 2011 ), or change the training set, while the test set keeps its original, real data (Rumi et al. 2018 ). Another dependency is the different kinds of aggregation that take place during modelling by time, space, or crime types attributes. For instance, while the majority of papers report to work with disaggregated crime types, some of them consider to aggregate crime types to, e.g., “violent crimes”, without specifying which types are included. In addition, the effects spatiotemporal aggregations have on the forecasting performance are typically not analysed, but could easily be conducted with a sensitivity analysis.

In this section, we perform a SWOT analysis of the most significant findings.

One of the strongest elements of current research efforts is the incorporation of spatial or spatiotemporal information into traditional prediction algorithms. The examples of this approach is STAR and STKDE (Shoesmith 2013 ; Rosser et al. 2017 ). Also, KDE, a traditional method in the field, has been adapted to consider sampling problems, such as sparse data (DCKDE) and grid format (NTKDE) (Malik et al. 2014 ; Rosser et al. 2017 ). Besides, the interest of the scientific community in the incorporation and effect of big data into prediction is evident from the related work section. This interest is also supported by the trend of introducing dynamic variables into the modelling process, such as calculating visitor entropy from Foursquare or ambient population from social networks and transportation. Regarding the performance evaluation, surveillance plots (Fig.  4 ) provide a more detailed picture of the accuracy of the forecasted information. Since they include the area coverage on the x-axis, they can be used by the police as a decision tool to identify the threshold that balances prediction accuracy with the size of patrolling areas.

Overall, significant details of study experiments are not always reported and commonly undefined items are the spatial unit of analysis and the sample size. Similarly, for feature-engineering dependent methods the crafting procedures are not sufficiently described. The above elements make a study difficult to reproduce or to compare its results with a possible future study. Furthermore, we did not find any open source tools that implement spatial crime forecasting using the best-proposed methods reported. Such a tool could enhance the possibility of reproducing results from an existing forecasting study. We suggest that all data items analysed in “ Overview of selected publications on spatial crime forecasting ” section (for an overview have a look at Table  3 ) should always be reported. However, a detailed “spatial forecasting protocol” could be developed similarly to protocols for other modelling approaches such as the ODD protocol (Grimm et al. 2010 ). Furthermore, the most common spatial unit is the grid cell, which may not necessarily align with places that policing resources are typically deployed to. So far, we did not encounter a study that sufficiently addresses this issue. Regarding the performance evaluation, most authors use standard metrics. A “global” standard metric, such as MAE, cannot describe the distribution of the prediction error across space, which can vary a lot. We thus propose to develop novel local spatial or spatiotemporal evaluation metrics. Finally, other modelling issues are hardly discussed, if at all, such as overfitting, multi-collinearity, sampling bias, and data sparsity.

Opportunities

There is a tremendous increase in spatial crime forecasting studies. From the pool of the 32 selected papers, 7 and 11 papers were published in 2017 and 2018, respectively, compared to about one paper per year between 2000 and 2016 (Fig.  2 ). This shows the growing interest of scholars from varying disciplines (compare Table  2 ) into this kind of research. The crime type that has been studied the most is residential burglary. It is unclear why this particular crime type and property crimes, in general, are more likely to be studied. A future opportunity could be to systematically test whether there is a pattern of property crimes to consistently outperforming other crime types and why. Furthermore, except for RTM and KDE, other spatial methods mentioned in the related work section (“ Related work ” section) were not used by the selected papers. The reason may be that authors have varying backgrounds, such as computer science and criminology, and may not be familiar with such methods. This opens a research opportunity to explore and compare less used spatial methods with traditional approaches, such as RTM or KDE. Another opportunity would be to compare the forecasting performance of the methods among each other. In this review, we presented methodological trends, but a fair comparison among spatial methods was not possible. First, some methods were not compared to a baseline method. Other authors compared the same method with a different set of features. Even if there were papers with a similar set of features a comparison among them would be biased due to variations of sample data, study areas, sampling periods, etc. Future empirical studies should focus on the comparison of algorithms, of which the number is constantly increasing. We merged the selected papers into four categories of forecasting algorithms, including the kernel-based, point processes, traditional machine learning, and deep learning. Traditional machine learning algorithms were present in most proposed methods, with MLP and RF being the most common ones, while AR models were the most used baselines methods. A suggestion is to compare new or recently developed algorithms to the most frequently proposed ones, instead of continuing to conduct further comparisons with traditional or simpler methods that have repeatedly shown to underperform.

We outlined that spatial crime forecasting studies lack coherent terminology, especially for terms such as “prediction”, “forecasting”, and “hotspots”. The predominant predictive task is the binary classification (n = 21) and the predominant forecasting inference is hotspots (n = 20). It is important to understand the rationale behind this trend. Is regression analysis less useful or more difficult to predict? Although we notice a constant increase in developing classification algorithms or features to be infused in the classification task, we acknowledge the importance of both prediction tasks. Also, for the display of an area’s crime picture, it is important to examine both hotspots and coldspots or a multiclass classification towards the hottest crime spots. However, none of these was the focus of the examined papers. We acknowledge that forecasting hotspots is important for police to allocate resources. Nevertheless, what about the information that can be derived by other types of spatial groupings such as coldspots, coldspot outliers, or hotspot outliers, commonly referred to as LL, LH, HL (low–low, low–high, high-low, respectively) and calculated by the local Moran statistic (Anselin 2005 )? Science needs to progress knowledge, which requires understanding and examining all aspects of a phenomenon. Finally, only a third of all papers performed long-term predictions. Although this trend is positive because law enforcement has an interest in almost real-time prediction, the long-term prediction should not be overlooked as playing an important role in the understanding of the crime risk and providing a broad picture for strategic planning.

In this paper, we focus on “Spatial Crime Forecasting”, which is an inference approach about crime both in time and in space. We conducted a systematic literature review that follows the reporting guidance “PRISMA” (Liberati et al. 2009 ) to understand and evaluate the state of the art concerning concepts and methods in empirical studies on crime with many applications and special attention to crime. We addressed several research questions that deal with the role of space in the forecasting procedure, the methods used, the predictive performance, and finally model validation strategies.

We identified five types of inference, namely (1) hotspots (the majority of the papers), (2) number of crime, (3) crime rate, (4) category of crime, (5) percent of crime in clusters, and (6), properties of clusters. With regards to forecasting methods, the authors proposed mostly traditional machine learning methods, but also kernel density estimation based approaches, and less frequently point process and deep learning approaches. When it comes to measuring performance, a plethora of metrics were used with the top three ones being 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.

This study was driven by the increasing publication of spatial crime forecasting studies and (crime predictive analytics in general). More than half of the selected papers (n = 32) were published in the last 2 years. In specific, about one paper per year was published between 2000 and 2016, while 7 and 11 papers were published in 2017 and 2018, respectively. At the same time, there is a global growth of scientific publication outputs. Bornmann and Mutz ( 2015 ), fitted an exponential model to this growth and calculated an increasing rate of outputs of about 3% annually, while the volume is estimated to double in approximately 24 years. Yet the yearly patterns of the selected papers show a much greater increase that indicates the importance and future potential of studies related to spatial crime forecasting.

Furthermore, we would like to outline the main limitations that may prohibit reproducibility, and hence the advancement of this topic in the long term. First, the terminology being used is not consistent possibly due to the fact that scientists working on this topic have various backgrounds (e.g. criminology, computer science, geosciences, public policy, etc.). Second, significant details of study experiments are vaguely or not at all reported. With respect to the last point, we suggested reporting the following data items: study area, scale, sampling period, months, type, sample, inference, task, spatial unit, and temporal unit (in total 10 items). Additional items to be reported are proposed method, best - proposed method, baseline method, evaluation metrics, and validation strategy (in total 5 items).

Availability of data and materials

The list of manuscripts used for this research is mentioned in Table  3 . If needed, the authors can provide the list of 193 manuscripts that went through the eligibility phase.

JCR: https://clarivate.com/webofsciencegroup/solutions/journal-citation-reports/ .

Al Boni, M., & Gerber, M. S. (2016). Predicting crime with routine activity patterns inferred from social media. In IEEE International Conference on Systems, Man and Cybernetics (SMC) , (pp. 1233–1238). https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7844410 .

Anselin, L. (2005). Exploring spatial data with GeoDaTM: A workbook . Santa Barbara: Center for Spatially Integrated Social Science.

Google Scholar  

Araújo, A., Cacho, N., Bezerra, L., Vieira, C., & Borges, J. (2018). Towards a crime hotspot detection framework for patrol planning. In 2018 IEEE 20th International Conference on High Performance Computing and Communications; IEEE 16th International Conference on Smart City; IEEE 4th International Conference on Data Science and Systems (HPCC/SmartCity/DSS) , (pp. 1256–1263). https://doi.org/10.1109/HPCC/SmartCity/DSS.2018.00211 .

Araujo, A. J., Cacho, N., Thome, A. C., Medeiros, A., & Borges, J. (2017). A predictive policing application to support patrol planning in smart cities. In International Smart Cities Conference (ISC2) . https://www.researchgate.net/profile/Adelson_Araujo2/publication/321236214_A_predictive_policing_application_to_support_patrol_planning_in_smart_cities/links/5c068339299bf169ae316a6f/A-predictive-policing-application-to-support-patrol-planning-in-smart-ci .

Bernasco, W., & Elffers, H. (2010). Statistical analysis of spatial crime data. In A. R. Piquero & D. Weisburd (Eds.), Handbook of quantitative criminology (pp. 699–724). New York: Springer. https://doi.org/10.1007/978-0-387-77650-7_33 .

Chapter   Google Scholar  

Bornmann, L., & Mutz, R. (2015). Growth Rates of Modern Science: A Bibliometric Analysis Based on the Number of Publications and Cited References. Journal of the Association for Information Science and Technology, 66 (11), 2215–2222.

Article   Google Scholar  

Bowen, D. A., Mercer Kollar, L. M., Wu, D. T., Fraser, D. A., Flood, C. E., Moore, J. C., Mays E. W. & Sumner, S. A. (2018). Ability of crime, demographic and business data to forecast areas of increased violence. International journal of injury control and safety promotion , 25 (4), 443–448. https://doi.org/10.1080/17457300.2018.1467461 .

Bramer, W. M., Rethlefsen, M. L., Kleijnen, J., & Franco, O. H. (2017). Optimal database combinations for literature searches in systematic reviews: A prospective exploratory study. Systematic Reviews, 6 (1), 245.

Brantingham, P. J., & Brantingham, P. L. (1984). Patterns in crime . New York: Macmillan.

Brayne, S. (2017). Big data surveillance: The case of policing. American Sociological Review, 82 (5), 977–1008.

Brown, D. E., & Oxford, R. B. (2001). Data mining time series with applications to crime analysis. In 2001 IEEE International Conference on Systems, Man and Cybernetics. e-Systems and e-Man for Cybernetics in Cyberspace (Cat. No. 01CH37236), Vol. 3 (pp. 1453–1458). IEEE. https://doi.org/10.1109/ICSMC.2001.973487 .

Bruinsma, G. J. N., & Johnson, S. D. (2018). The oxford handbook of environmental criminology . Oxford: Oxford University Press.

Book   Google Scholar  

Caplan, J. M., Kennedy, L. W., & Miller, J. (2011). Risk terrain modeling: brokering criminological theory and gis methods for crime forecasting. Justice Quarterly, 28 (2), 360–381. https://doi.org/10.1080/07418825.2010.486037 .

Chainey, S., Tompson, L., & Uhlig, S. (2008). The utility of hotspot mapping for predicting spatial patterns of crime. Security Journal., 21, 4–28.

Chauhan, C., & Sehgal, S. 2017. A review: crime analysis using data mining techniques and algorithms. In P. N. Astya, A. Swaroop, V. Sharma, M. Singh, & K Gupta, (Ed.), 2017 IEEE International Conference on Computing, Communication and Automation (ICCCA) , edited by , (pp. 21–25).

Chen, H. C., Chung, W. Y., Xu, J. J., Wang, G., Qin, Y., & Chau, M. (2004). Crime data mining: A general framework and some examples. Computer, 37 (4), 50–56. https://doi.org/10.1109/MC.2004.1297301 .

Cohen, J., Gorr, W. L., & Olligschlaeger, A. M. (2007). Leading indicators and spatial interactions: A crime‐forecasting model for proactive police deployment. Geographical Analysis , 39 (1), 105–127. https://doi.org/10.1111/j.1538-4632.2006.00697.x

Cressie, N. A. C. (1993). Statistics for spatial data ., New York: Wiley. https://doi.org/10.2307/2533238 .

Dash, S. K., Safro, I., & Srinivasamurthy, R. S. (2018). Spatio-temporal prediction of crimes using network analytic approach. In 2018 IEEE International Conference on Big Data (Big Data) (pp. 1912-1917). IEEE. https://doi.org/10.1109/BigData.2018.8622041 .

Drawve, G., Moak, S. C., & Berthelot, E. R. (2016). Predictability of gun crimes: A comparison of hot spot and risk terrain modelling techniques. Policing & Society, 26 (3), 312–331. https://doi.org/10.1080/10439463.2014.942851 .

Dugato, M., Favarin, S., & Bosisio, A. (2018). Isolating target and neighbourhood vulnerabilities in crime forecasting. European of Criminal Policy and Reserach., 24 (4 SI), 393–415. https://doi.org/10.1007/s10610-018-9385-2 .

Gerber, M. S. (2014). Predicting crime using twitter and kernel density estimation. Decision Support Systems., 61, 115–125. https://doi.org/10.1016/j.dss.2014.02.003 .

Gimenez-Santana, A., Caplan, J. M., & Drawve, G. (2018). Risk terrain modeling and socio-economic stratification: Identifying risky places for violent crime victimization in Bogota, Colombia. European of Criminal Policy and Reserach, 24 (4 SI), 417–431. https://doi.org/10.1007/s10610-018-9374-5 .

Gorr, W. L. (2009). Forecast accuracy measures for exception reporting using receiver operating characteristic curves. International Journal of Forecasting, 25 (1), 48–61.

Gorr, W., & Harries, R. (2003). Introduction to crime forecasting. International Journal of Forcasting 19. https://www.sciencedirect.com/science/article/pii/S016920700300089X .

Gorr, W., Olligschlaeger, A., & Thompson, Y. International Journal Of, and Undefined 2003. (2003). “Short-Term Forecasting of Crime.” International Journal of Forecasting . https://www.sciencedirect.com/science/article/pii/S016920700300092X .

Grimm, V., Berger, U., DeAngelis, D. L., Polhill, J. G., Giske, J., & Railsback, S. F. (2010). The ODD protocol: A review and first update. Ecological Modelling, 221 (23), 2760–2768.

Haddaway, N. R., Collins, A. M., Coughlin, D., & Kirk, S. (2015). The role of google scholar in evidence reviews and its applicability to grey literature searching. PloS ONE 10(9).

Hardyns, W., & Rummens, A. (2018). predictive policing as a new tool for law enforcement? Recent developments and challenges. European Journal on Criminal Policy and Research, 24 (3), 201–218. https://doi.org/10.1007/s10610-017-9361-2 .

Hart, T., & Zandbergen, P. (2014). Kernel density estimation and hotspot mapping examining the influence of interpolation method, grid cell size, and bandwidth on crime forecasting. Policing—An International Journal o FPolice Strategies & Management, 37 (2), 305–323. https://doi.org/10.1108/PIJPSM-04-2013-0039 .

Hassani, H., Huang, X., Silva, E. S., & Ghodsi, M. (2016). A review of data mining applications in crime. Statistical Analysis and Data Mining, 9 (3), 139–154. https://doi.org/10.1002/sam.11312 .

Holone, H. (2016). The filter bubble and its effect on online personal health information. Croatian Medical Journal, 57 (3), 298.

Hu, Y., Wang, F., Guin, C., Zhu, H. (2018). A spatio-temporal Kernel density estimation framework for predictive crime hotspot mapping and evaluation. Applied Geography 99:89–97. https://www.sciencedirect.com/science/article/pii/S0143622818300560 .

Huang, C., Zhang, J., Zheng, Y., & Chawla, N. V. (2018). DeepCrime: Attentive hierarchical recurrent networks for crime prediction. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management , (pp. 1423–1432). CIKM’18. New York, NY, USA: ACM. https://doi.org/10.1145/3269206.3271793 .

Hunt, J. M. (2016). Do crime hot spots move? Exploring the effects of the modifiable areal unit problem and modifiable temporal unit problem on crime hot spot stability. American University.

Ivaha, C., Al-Madfai, H., Higgs, G., & Ware, J. A. (2007). The dynamic spatial disaggregation approach: A spatio-temporal modelling of crime. In World Congress on Engineering , (pp. 961–966). Lecture Notes in Engineering and Computer Science. http://www.iaeng.org/publication/WCE2007/WCE2007_pp961-966.pdf .

Johansson, E., Gåhlin, C., & Borg, A. (2015). Crime hotspots: An evaluation of the KDE spatial mapping technique. In 2015 European Intelligence and Security Informatics Conference , (pp. 69–74). https://doi.org/10.1109/EISIC.2015.22 .

Kadar, C., Brüngger, R. R., & Pletikosa, I. (2017). Measuring ambient population from location-based social networks to describe urban crime. In International Conference on Social Informatics , (pp. 521–35). Springer, New York.

Kadar, C., & Pletikosa, I. (2018). Mining large-scale human mobility data for long-term crime prediction. EPJ Data Science . https://doi.org/10.1140/epjds/s13688-018-0150-z .

Kennedy, L. W., & Caplan, J. M. (2012). A theory of risky places . Newark: Rutgers Center on Public Security.

Kennedy, L. W., & Dugato, M. (2018). “Forecasting crime and understanding its causes. Applying risk terrain modeling worldwide. European Journal on Criminal Policy and Research, 24 (4, SI), 345–350. https://doi.org/10.1007/s10610-018-9404-3 .

Kinney, J. B., Brantingham, P. L., Wuschke, K., Kirk, M. G., & Brantingham, P. J. (2008). Crime attractors, generators and detractors: Land use and urban crime opportunities. Built Environment, 34 (1), 62–74.

Liberati, A., Altman, D., Tetzlaff, J., Mulrow, C., Gøtzsche, P. C., Ioannidis, J. P. A., et al. (2009). The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: Explanation and elaboration. Journal of Clinical Epidemiology . https://doi.org/10.1016/j.jclinepi.2009.06.006 .

Liesenfeld, R., Richard, J. F., & Vogler, J. (2017). Likelihood-based inference and prediction in spatio-temporal panel count models for urban crimes. Journal of Applied Econometrics, 32 (3), 600–620. https://doi.org/10.1002/jae.2534 .

Lin, Y. L., Yen, M. F., & Yu, L. C. (2018). Grid-based crime prediction using geographical features. ISPRS International Journal of Geo-Information, 7 (8), 298.

Malik, A., Maciejewski, R., Towers, S., McCullough, S., & Ebert, D. S. (2014). Proactive spatiotemporal resource allocation and predictive visual analytics for community policing and law enforcement. IEEE Transactions on Visualization and Computer 20(12): 1863–72. https://www.computer.org/csdl/trans/tg/2014/12/06875970-abs.html .

Mohler, G. (2014). Marked point process hotspot maps for homicide and gun crime prediction in Chicago. International Journal of Forecasting, 30 (3), 491–497. https://doi.org/10.1016/j.ijforecast.2014.01.004 .

Mohler, G., & Porter, M. D. (2018). Rotational grid, PAI-maximizing crime forecasts. Statistical Analysis and Data Mining: The ASA Data Science Journal , 11 (5), 227-236. https://doi.org/10.1002/sam.11389 .

Mohler, G., Raje, R., Carter, J., Valasik, M., & Brantingham, J. (2018). A penalized likelihood method for balancing accuracy and fairness in predictive policing. In 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC) . https://ieeexplore.ieee.org/abstract/document/8616417/ .

Mohler, G. O., Short, M. B., Brantingham, P. J., Schoenberg, F. P., & Tita, G. E. (2011). Self-exciting point process modeling of crime. Journal of the American Statistical Association, 106 (493), 100–108.

Mohler, G. O., Short, M. B., Malinowski, S., Johnson, M., Tita, G. E., Bertozzi, A. L., et al. (2015). Randomized controlled field trials of predictive policing. Journal of the American Statistical Association, 110 (512), 1399–1411. https://doi.org/10.1080/01621459.2015.1077710 .

Mu, Y., Ding, W., Morabito, M., & Tao, D. (2011). Empirical discriminative tensor analysis for crime forecasting. In International Conference on Knowledge Science, Engineering and Management (pp. 293-304). Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-25975-3_26 .

Ohyama, T., & Amemiya, M. (2018). applying crime prediction techniques to Japan: A comparison between risk terrain modeling and other methods. European Journal on Criminal Policy and Research, 24 (4), 469–487.

Ozkan, T. (2018). Criminology in the age of data explosion: new directions. The Social Science Journal . https://doi.org/10.1016/J.SOSCIJ.2018.10.010 .

Papamitsiou, Z., & Economides, A. A. (2014). Learning analytics and educational data mining in practice: A systematic literature review of empirical evidence. Journal of Educational Technology & Society, 17 (4), 49–64.

Perry, W. L. (2013). Predictive policing: The role of crime forecasting in law enforcement operations . Santa Monica: Rand Corporation.

Ratcliffe, J. (2015). What is the future… of predictive policing. Practice, 6 (2), 151–166.

Rodríguez, C. D., Gomez, D. M., & Rey, M. A. (2017). Forecasting time series from clustering by a memetic differential fuzzy approach: An application to crime prediction. In 2017 IEEE Symposium Series on Computational Intelligence (SSCI) , (pp. 3372–3379). https://ieeexplore.ieee.org/abstract/document/8285373 .

Rosser, G., Davies, T., Bowers, K. J., Johnson, D. S., & Cheng, T. (2017). Predictive crime mapping: Arbitrary grids or street networks? Journal of Quantitative Criminology, 33 (3), 569–594. https://doi.org/10.1007/s10940-016-9321-x .

Rumi, S. K., Deng, K., & Salim, F. D. EPJ Data Science, and Undefined 2018. (2018). Crime event prediction with dynamic features. EPJ Data Science . https://doi.org/10.1140/epjds/s13688-018-0171-7 .

Rummens, A., Hardyns, W., & Pauwels, L. (2017). The use of predictive analysis in spatiotemporal crime forecasting: Building and testing a model in an urban context. Applied Geography, 86, 255–261. https://doi.org/10.1016/j.apgeog.2017.06.011 .

Seele, P. (2017). Predictive sustainability control: A review assessing the potential to transfer big data driven ‘predictive policing’ to corporate sustainability management. Journal of Cleaner Production, 153, 73–86. https://doi.org/10.1016/j.jclepro.2016.10.175 .

Shamsuddin, N.H. M., Ali. N. A., & Alwee, R. (2017). An overview on crime prediction methods. In 6th ICT International Student Project Conference (ICT - ISPC), IEEE. https://ieeexplore.ieee.org/abstract/document/8075335/ .

Shoesmith, G. L. (2013). Space–time autoregressive models and forecasting national, regional and state crime rates. International Journal of Forecasting, 29 (1), 191–201. https://doi.org/10.1016/j.ijforecast.2012.08.002 .

Thongsatapornwatana, U. (2016). A survey of data mining techniques for analyzing crime patterns. In 2016 Second Asian Conference on Defence Technology (ACDT) , (pp. 123–28). https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7437655 .

Thongtae, P., & Srisuk, S. (2008). An analysis of data mining applications in crime domain. In X. He, Q. Wu, Q. V. Nguyen, & W. Ja (Ed.), 8th IEEE International Conference on Computer and Information Technology Workshops , (pp. 122–126). https://doi.org/10.1109/CIT.2008.Workshops.80 .

Vlahogianni, E. I., Karlaftis, M. G., & Golias, J. C. (2014). Short-term traffic forecasting: where we are and where we’re going. Transportation Research Part C: Emerging Technologies, 43, 3–19.

Wang, X., & Brown, D. E. (2011). The Spatio-Temporal Generalized Additive Model for Criminal Incidents. In Proceedings of 2011 IEEE International Conference on Intelligence and Security Informatics , (pp. 42–47). IEEE, New York.

Wang, X., Brown, D. E., Gerber, M.S. (2012). Spatio-temporal modeling of criminal incidents using geographic, demographic, and twitter-derived information. In 2012 IEEE International Conference on Intelligence and Security Informatics , (pp. 36–41). IEEE, New York.

Wang, H., Kifer, D., Graif, C., & Li, Z. (2016). Crime rate inference with big data. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , (pp. 635–644). KDD’16. New York, NY, USA: ACM. https://doi.org/10.1145/2939672.2939736 .

Williams, M. L., & Burnap, P. (2015). Cyberhate on social media in the aftermath of woolwich: A case study in computational criminology and big Data. British Journal of Criminology, 56 (2), 211–238.

Yang, D., Heaney, T., Tonon, A., Wang, L., & Cudré-Mauroux, P. (2018). CrimeTelescope: crime hotspot prediction based on urban and social media data fusion. World Wide Web , 21 (5), 1323–1347. https://doi.org/10.1007/s11280-017-0515-4 .

Yu, C. H., Ward, M. W., Morabito, M., & Ding, W. (2011). Crime forecasting using data mining techniques. In IEEE 11th International Conference on Data Mining Workshops , (pp. 779–786). IEEE, New York.

Zhao, X., & Tang, J. (2017). Modeling temporal-spatial correlations for crime prediction. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management , (pp. 497–506). CIKM’17. New York, NY, USA: ACM. https://doi.org/10.1145/3132847.3133024 .

Zhuang, Y., Almeida, M., Morabito, M., & Ding. W. (2017). Crime hot spot forecasting: A recurrent model with spatial and temporal information. In X. D, Wu, T. Ozsu, J. Hendler, R. Lu, (Ed.), IEEE International Conference on Big Knowledge (ICBK) , (pp. 143–150). https://doi.org/10.1109/ICBK.2017.3 .

Download references

This research was funded by the Austrian Science Fund (FWF) through the Doctoral College GIScience at the University of Salzburg (DK W 1237-N23).

Author information

Authors and affiliations.

Department of Geoinformation Processing, University of Twente, Enschede, The Netherlands

Ourania Kounadi

Doctoral College GIScience, Department of Geoinformatics-Z_GIS, University of Salzburg, Salzburg, Austria

Alina Ristea & Michael Leitner

Boston Area Research Initiative, School of Public Policy and Urban Affairs, Northeastern University, Boston, MA, USA

Alina Ristea

Department of Informatics and Applied Mathematics, Federal University of Rio Grande do Norte, Natal, RN, Brazil

Adelson Araujo Jr.

Department of Geography and Anthropology, Louisiana State University, Baton Rouge, LA, USA

Michael Leitner

You can also search for this author in PubMed   Google Scholar

Contributions

OK designed the initial idea, conceived and designed all experiments, analyzed the data and wrote the paper. AR helped designing the paper and running experiments. She also helped writing parts of the paper, with a focus on forecasting performance. AAJr. gave technical support and helped with running experiments and conceptual framework. He also helped writing the manuscript, focusing on forecasting algorithms. ML supervised the work and edited the final manuscript. All authors discussed the results and implications and commented on the manuscript at all stages. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Alina Ristea .

Ethics declarations

Competing interests.

The authors declare that they have no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Additional file 1..

Online survey on Risk of Bias across Studies.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Kounadi, O., Ristea, A., Araujo, A. et al. A systematic review on spatial crime forecasting. Crime Sci 9 , 7 (2020). https://doi.org/10.1186/s40163-020-00116-7

Download citation

Received : 17 November 2019

Accepted : 11 May 2020

Published : 27 May 2020

DOI : https://doi.org/10.1186/s40163-020-00116-7

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Forecasting
• Predictive policing
• Spatiotemporal
• Spatial analysis

Crime Science

ISSN: 2193-7680

• General enquiries: [email protected]

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Subscribe to the PwC Newsletter

Join the community, edit social preview.

Add a new code entry for this paper

Remove a code repository from this paper, mark the official implementation from paper authors, add a new evaluation result row.

• CRIME PREDICTION

Remove a task

Add a method, remove a method, edit datasets, crime prediction using machine learning and deep learning: a systematic review and future directions.

28 Mar 2023  ·  Varun Mandalapu , Lavanya Elluri , Piyush Vyas , Nirmalya Roy · Edit social preview

Predicting crime using machine learning and deep learning techniques has gained considerable attention from researchers in recent years, focusing on identifying patterns and trends in crime occurrences. This review paper examines over 150 articles to explore the various machine learning and deep learning algorithms applied to predict crime. The study provides access to the datasets used for crime prediction by researchers and analyzes prominent approaches applied in machine learning and deep learning algorithms to predict crime, offering insights into different trends and factors related to criminal activities. Additionally, the paper highlights potential gaps and future directions that can enhance the accuracy of crime prediction. Finally, the comprehensive overview of research discussed in this paper on crime prediction using machine learning and deep learning approaches serves as a valuable reference for researchers in this field. By gaining a deeper understanding of crime prediction techniques, law enforcement agencies can develop strategies to prevent and respond to criminal activities more effectively.

Code Edit Add Remove Mark official

Tasks edit add remove, datasets edit, results from the paper edit, methods edit add remove.

• Architecture and Design
• Asian and Pacific Studies
• Business and Economics
• Classical and Ancient Near Eastern Studies
• Computer Sciences
• Cultural Studies
• Engineering
• General Interest
• Geosciences
• Industrial Chemistry
• Islamic and Middle Eastern Studies
• Jewish Studies
• Library and Information Science, Book Studies
• Life Sciences
• Linguistics and Semiotics
• Literary Studies
• Materials Sciences
• Mathematics
• Social Sciences
• Sports and Recreation
• Theology and Religion
• Publish your article
• The role of authors
• Promoting your article
• Abstracting & indexing
• Publishing Ethics
• Why publish with De Gruyter
• How to publish with De Gruyter
• Our book series
• Our subject areas
• Your digital product at De Gruyter
• Contribute to our reference works
• Product information
• Tools & resources
• Product Information
• Promotional Materials
• Orders and Inquiries
• FAQ for Library Suppliers and Book Sellers
• Repository Policy
• Free access policy
• Open Access agreements
• Database portals
• For Authors
• Customer service
• People + Culture
• Journal Management
• How to join us
• Working at De Gruyter
• Mission & Vision
• De Gruyter Foundation
• De Gruyter Ebound
• Our Responsibility
• Partner publishers

Your purchase has been completed. Your documents are now available to view.

A study on predicting crime rates through machine learning and data mining using text

Crime is a threat to any nation’s security administration and jurisdiction. Therefore, crime analysis becomes increasingly important because it assigns the time and place based on the collected spatial and temporal data. However, old techniques, such as paperwork, investigative judges, and statistical analysis, are not efficient enough to predict the accurate time and location where the crime had taken place. But when machine learning and data mining methods were deployed in crime analysis, crime analysis and predication accuracy increased dramatically. In this study, various types of criminal analysis and prediction using several machine learning and data mining techniques, based on the percentage of an accuracy measure of the previous work, are surveyed and introduced, with the aim of producing a concise review of using these algorithms in crime prediction. It is expected that this review study will be helpful for presenting such techniques to crime researchers in addition to supporting future research to develop these techniques for crime analysis by presenting some crime definition, prediction systems challenges and classifications with a comparative study. It was proved though literature, that supervised learning approaches were used in more studies for crime prediction than other approaches, and Logistic Regression is the most powerful method in predicting crime.

1 Introduction

Violations of the law pose a danger to the administration of justice and should be curtailed. Computational crime prediction and forecasting can help improve the safety of metropolitan areas. The inability of humans to process large amounts of complicated data from big data makes it difficult to make early and accurate predictions about criminal activity. Computational problems and opportunities arise from accurately predicting crime rates, types, and hot locations based on historical patterns. Still, there is a need for stronger prediction algorithms that target police patrols toward criminal events, despite extensive research efforts [ 1 ].

Crime analysis is a methodology approach used to identify crime spots and it is not an easy approach. In year 2020, Geographical Information Systems (GIS) was the non-machine learning tool used earlier for temporal and spatial data. GIS used the crime spots technique that mainly depends on crime type to help reduce crime rates [ 2 ].

Crime rate prediction can be defined as a method to build a system for finding crime future patterns and help the law enforcer to solve the crime which lead to reduce its rate in the real-world. Meanwhile, crime forecasting refers to the ability to predict far future crimes, up to years in the future to increase crime preventions, and this can be achieved by using time series approaches to find future crime trends from time series data.

In general, crime analysis in data mining can be predicted using different methods such as statistical methods [ 3 , 4 , 5 ], cover visualization methods [ 6 , 7 , 8 ], unsupervised learning, and supervised learning techniques [ 9 , 10 , 11 ]. Visualization methods include visual explanation of the connection between geographical view and other crime data such as geographic profiling [ 12 ], GIS-based crime mapping [ 13 ], crime prediction [ 14 , 15 , 16 ], and asymmetric mapping [ 17 ]. However, to obtain the connection between statistical methods, unsupervised learning techniques and crime data such as clustering methods which were very popular. These techniques were implemented as follows: clustering methods were used as criminal behavior analysis [ 18 , 19 ], crime pattern recognition, criminal association analysis, and incident pattern recognition to extract the groups or patterns that had the same features in crime data [ 20 ].

Then, the machine learning algorithms’ development helped the crime data analysis researchers to investigate crime depending on preprocessing and clustering techniques to extract the crime locations from row data [ 21 ], using the supervised and unsupervised machine learning models to analyze these data and discover their pattern based on time and location of crime to produce precise predictions [ 22 ]. In addition, the machine learning algorithms’ development also helped to investigate the reasons of crime occurring in certain areas by applying machine learning algorithms on history data collected from past years in the same area [ 23 ].

Nowadays, the development of classification algorithms, especially machine learning algorithms, helps to enhance the crime prediction [ 24 ]. Therefore, researchers tried to connect crime with time depending on various factors to help in resolving the crimes and prevent it and its frequencies. In year 2018, Fourier series was proposed as an analytic technique to accomplish a flexible mathematical model on time periodic effects. This technique explained the accuracy and usefulness of analytical techniques to connect the time factor with crime prediction. Thereby, the analytical techniques effectively achieved the relation between crimes and time, but not for all type of crimes [ 25 ].

We can say that, machine learning algorithms is widely used in crime prediction discipline, but it is not more than data mining and each one has its own performance and gives a perfect result.

Our work has been setup so that interested parties become familiar with the previous studies and the accuracy that have been achieved, presented in tabular format. The main contribution of this study is presenting machine learning and data mining applications in predicting crimes, by classifying the studies according to different types of techniques, and providing a brief overview of each applicable methodology that has been used to mine crime, and also, enlisting some challenges faced by such system developers.

The limitation of the state of art works are the lack with big entered geo-area, no generality because of using the same system on two different crime datasets leads to different accuracy percentages with big difference, the lack of works that predict criminal action, and finally but not last, the difficulty that faced the researchers in the crime prediction field, that it may be a missing informations in the on-line crime datasets or the data are repeated.

The rest of the review paper is organized as follows: in Section 2 , the research methodology of the survey is explained, in Section 3 , crime definitions and descriptions are discussed in detail. In Section 4 , challenges of prediction system are discussed. In Section 5 , the public datasets are described. The related work is included in Section 6 . In Section 7 , the prediction system classification is introduced. A comparison study of previous works is explained in Sections 8.1 and 8.2 . Eventually, discussion and conclusion are presented at the end of this article in Section 9 .

2 Research methodology

The methodology involved in this review study contains two stages: first is getting the relevant research works on crime prediction with machine learning and data mining studies and analyze them, and second is setting a classification table in Section 8 , and finally presenting a study about the performances of various algorithms and the achieved accuracy and comparison between them.

In choosing relevant research works, any Master and Doctoral dissertations or any papers that were not published were ignored. The research keyword was crime prediction with machine learning and data mining or violent crime prediction, the publishing criteria was between 2001 and 2022, the abstract of every article was read and then determined if it is relevant or not.

3 Crime definition and description

Generally, crimes are classified in to three groups: infractions, felonies, and misdemeanors based on the severity, punishment, and seriousness of crimes. Infractions are minimal crimes such as tailgating, parking overtime, and speeding. Meanwhile, Felonies are considered as most severe crimes followed by misdemeanors which are considered less severe crimes [ 26 ]. In addition, the crimes are classified into types based on the time when occurred such as the day, week, month, and season in order to find the connection between these types of crime and then to predict them in the future using machine learning and data mining algorithms. This can be done by using a dataset collected on a certain area for earlier crimes to forecast the future ones.

There are many types of crimes depending on the severity of the crime. Therefore, crimes are classified into three types, which are, felony, misdemeanors, and infractions (or wobblers) [ 22 ], as listed and defined in Table 1 .

Crime description

In addition, a crime could be categorized in other categories, such as victim, victimless, and violent crimes and there are other categorizations for crime, but through this study, only the classification mentioned in Table 1 will be considered.

4 Prediction systems challenges

Researchers and governmental security agents face some problems when it comes to predict crime’s location, time, and problems in choosing the effective method to do so. In addition, there are problems faced by the computer science researchers who used machine learning, data mining, and spatial–temporal data. In 2012 and 2016, the near-repeat-victimization and repeat-victimization methods were implemented to predict crimes in houses, streets, and regions. These methods state that if a crime happened in a block, then there is a probability that other crimes are increasing significantly in the same area [ 27 , 28 ].

The huge amount of data requires a large amount of storage

Crime-related data are usually in different formats such as text, images, graph, audio, relational data, unstructured data, and semi-structured data [ 29 ], so, the process of transforming these data to the understandable format is also a challenge.

In machine learning, to give the correct label (e.g., prediction or output) to an instance (e.g., context or input) is a challenge.

Use of appropriate data mining algorithm that gives better results than the used algorithms.

The environment and surrounding factors, such as the lack of the law and the weather, have an impact on the likelihood of crime, which ultimately causes the crime prediction algorithms to make grave errors. Any crime forecast must take the surrounding and environmental changes into consideration to avoid making such errors and to achieve high prediction accuracy.

5 Crime datasets

Crime-related data are gathered from a variety of different sources, including police reports, social media, news, and criminal records. It is difficult to gather data of this amount [ 30 ]. The datasets are available online in many countries around the world or gathered from the police departments. During our survey, we noticed that the Chicago crime dataset is more frequently used in crime prediction systems, and that returned to the large population and hight crime rates in this area [ 31 ].

6 Related work

With the huge data size nowadays, the evaluation of machine learning and data mining techniques allow us to deal with this row data and extract the results in better ways. Techniques for criminal activity detection and, more generally, machine learning and data mining, have recently been applied to the area of policing to achieve crime reduction.

Correct choices of the parameters for these techniques can help law enforcers to analyze and find the likelihood between crimes as well as patterns and trends in criminal activities, which lead to qualify those activities more efficiently [ 5 ].

In this section, the previous related works are discussed and analyzed, these research works are widely variate, some of these take the field of crime analysis to predict, some take the field of application of Artificial Intelligent on crime data, machine learning or data mining (which are subfields of Artificial Intelligent) in order to predict and forecast violence crimes, based on spatial and temporal data in some research works.

During our survey, we noticed five surveys or overviews related to crime prediction and machine learning or data mining.

The earliest was in 2011, a survey of different methods that used to extract patterns from spatial information (they called it spatial data mining (SDM) algorithms) like co-location mining, spatial clustering, spatial hot spots, spatial outliers, spatial auto-regression, conditional auto-regression, and geographically weighted regression, which conclude the effectiveness of these SDM algorithms and the guarantee to use it in the real world, and they found the need for more methods to validate the hypotheses produced by these algorithms [ 32 ].

In 2015, some researches in the field of crime prediction with data mining and machine learning were discussed , this research takes a variety of crime related variables then found that the information influencing the crime rate such as age, alcohol, hot spots, media, some policies, etc., do not have effect on crime rate prediction [ 33 ], it succeeded in discussion, but there is a shortage in the conclusion.

In 2016, another survey was published. It reviewed over 100 applications of data mining in crime. They made a concise review by preparing a brief table containing the used technique with a specific software, the relevant study area with the expected use and function. They suggested to enhance the benefits, improvements, and usability of data mining techniques in crime data mining by introducing more training and educating fields for these techniques [ 34 ].

In 2019, a systematic review of crime prediction and data mining studies between 2004 and 2018 classified the research works based on the used data mining techniques. Based on the challenges addressed and the number of research papers according to technique used, by covering 40 papers, a gap was identified in all of them, that is, when datasets increase, there is a noticeable decrease in the system’s overall performance [ 35 ].

Finally, in 2020, another systematic review was done, 32 papers were analyzed from 2000 to 2018 in spatial crime forecasting. In this study, in addition to the surveying table that contains the information about the space and time of the research, the crime data, and forecasting details, more than one summary was given, that is, the top four proposed methods, best proposed, and baseline methods applied in the 32 selected papers. This study discussed the points of strengths, weaknesses, threats, and opportunities of the selected papers, and the conclusion was that the contiguity of algorithms should not be ignored in the future [ 2 ].

7 Classification of prediction systems

According to approaches, machine learning and data mining.

According to prediction type, special and temporal.

According to dataset, image prediction and data prediction.

8 Comparison study: Crime prediction vs classification approaches

In this section, Tables 2 and 3 lists the literature surveys of the machine learning and data mining algorithms using different datasets for different cities around the world. In addition, a comparison is made between machine learning and data mining methods toward crime in a border crime prediction system. In these tables, we enlist each selected paper with the important information that will assist other researchers in determining which categories of crime prediction techniques are most powerful. Consequently, these two tables explain the machine learning and data mining algorithms with crime prediction in order to achieve the purpose of this survey. The tables contain the references, the machine learning or data mining algorithms, the used dataset source, and the accuracy of each algorithm depending on a certain dataset that was used for a particular city. The following section discusses crime prediction research works that followed the machine learning and data mining approaches, separately.

Literature survey on crime prediction research works with machine learning

Literature survey on crime prediction research works with data mining

8.1 Machine learning and crime prediction

Crime prediction has been studied widely due to its relation with the society, these studies employ machine learning algorithms to outfit the crime predicting and forecasting issues. Machine learning algorithms are successfully used to predict spatial crime information. So, in 2006, Support Vector Machine (SVM) algorithm was applied to predict the location of crimes in Columbus, Ohio, US. SVM used both random and clustering approaches to train and test dataset and then predict the hot spot area and improve its effectiveness [ 37 ]. These algorithms are used to study the correlation between crime occurrence and crime motivates. In 2013, a Logistic Regression (LR) algorithm was implemented to forecast the relationship between burglar crimes and several other factors which are time of the day, day of the week, barriers, connectors, and repeated victimization, but this model was a failure for large geo-area [ 38 ].

In 2015, crime was predicted in southern US states using Random Forest (RF) method after applying SmoteR algorithm to detect the more dangerous crimes. In addition, their work was optimized using R software after the density and population were selected as real values [ 39 ].

Eventually, the auto-regressive approach was implemented to forecast the number of crimes that happened in the same time and predict them in urban areas [ 40 ]. In 2017, Naive Bayes (NB) algorithm was proposed to predict crime incident depending on history data that shows the same crime happening in the same place. Moreover, NB model was compared with Decision Tree (DT) algorithm in order to test the performance of the proposed method, and found that the NB outperforms the DT even with the computational complexity of DT [ 41 ].

In 2020, many research works were presented, one of them fused three methods, the Long short-term memory (LSTM), Residual neural network, and Graph convolutional network to propose a certain mechanism, which was able to extract spatial–temporal features to predict crimes in Chicago. In addition, Root mean square error and Mean absolute error were used as a criterion to test the performance of the applied method [ 42 ]. On the other study, a crime network for spatiotemporal data was proposed using Convolutional neural network (CNN) in order to automatically predict the time and place of the crimes [ 43 ]. And in another study [ 44 ], Recurrent neural network (RNN) with LSTM was integrated in order to design time series crime prediction system to predict crimes in Addis Ababa. Also, in one more study [ 45 ], the severity level of crime in Boston was studied and predicted using machine learning algorithm such as SVM, NB, LR, and DT.

According to ref. [ 31 ], the Deep neural network (DNN) has overcome the SVM, but according to ref. [ 46 ], the opposite occurred, the SVM has overcome the DNN, and this can be justified by one reason, the first has worked on an image dataset and the second has worked on a text dataset. So, it is recommended to use DNN in case of an image crime dataset.

According to refs [ 1 , 47 ], using the same system on two different crime datasets leads to different accuracy percentages with big difference, which shows that the dataset utilized severely affects the results gained. Therefore, this presents a challenge to these algorithms to prove its efficiency and then its accuracy to predict a crime.

After surveying the machine learning approaches, the highest accuracy crime prediction results gained are shown in Table 2 .

According to ref. [ 48 ], the LR algorithm achieves the highest accuracy among the different machine learning algorithms.

When observing the crime prediction results of the research works adopting the RF method, it was noticed that the highest accuracy achieved is 59.8%, which is considered a poor accuracy, compared with other methods.

The standard deviations of crime prediction accuracies for each algorithm show that the SVM algorithm outperforms the LR algorithm and achieves (71.9%) accuracy. Actually, it outperforms all other machine learning algorithm’s standard deviation results.

According to the previous studies, it was noticed that the highest crime prediction accuracy results were gained through the machine learning logistic regression method, which was 95% for Baltimore city in ref. [ 48 ]. Furthermore, algorithms such as XGBoost and Logistic Regression have achieved a high accuracy of 94 and 90%, respectively [ 1 ]. However, it can be noticed that the same algorithm can perform differently with two different datasets, and this proves that the dataset has a large influence on the crime prediction results.

8.2 Data mining and crime prediction

In 2011, special data mining and technologies were proposed to extract patterns from spatial and temporal data. In addition, the data were mined geospatially using special knowledge. In 2011, crimes were predicted in Portland; data mining methods were used to forecast crimes using spatial and temporal dataset collected in Portland and predict whether residential burglary will happen. The methods NB, SVM, DT, and K-Nearest Neighbor algorithms were applied to predict crimes and the result was compared between these methods, which shows the power of neural network in complex systems [ 57 ]. Moreover, the pattern extraction usefulness was limited by the complexity of the relationships between spatial data [ 32 ]. In 2016, high accuracy was achieved using various DT algorithms to extract knowledge from data collected during 1994 instances, with 128 attributes, then made a comparison between them. In addition, the data were trained and tested using scatter plots to illustrate the crime areas with the severity of each area based on previous data [ 58 ]. In year 2016, data mining algorithms were developed and used to classify these crimes based on their types. A crime was characterized according to time, based on factors such as vacations that started with the academic year for colleges and schools. In addition, the classifier was used to predict the severity risk of the crime areas in Denver city between 2010 and 2015 [ 59 ]. In 2020, Autoregressive Integrated Moving Average (ARIMA) technique was implemented to predict time series data and then have been visualized with data mining platform. This technique proved that regressive model can work on historical newsfeed data to predict future crimes [ 60 ].

Table 3 shows the comparison of many algorithms implemented against crime prediction challenge, such as DT, NB, RF, etc., either individually or group of them to a certain type of dataset and city. Thereby, this presents a challenge to these algorithms to confirm its effectiveness and then its accuracy to predict a crime.

The highest accuracy crime prediction results gained, based on the survey of the data mining methods, are shown in Table 3 .

According to ref. [ 61 ] the K-mean algorithm achieves the highest accuracy among the different data mining algorithms.

When we take the standard deviations of crime prediction accuracies for each algorithm, we noticed that the DT algorithm outperforms the NB algorithm and achieves (18.9%) accuracy.

According to the previous study, DT and Neural Network have recorded 94% accuracy for different datasets in refs [ 48 , 51 ] for machine learning algorithms. The k-mean data mining algorithm achieved 93.62% (cluster one) and 93.99% (cluster two) for crimes in India [ 61 ].

9 Conclusion

Crime prediction became the hot research area nowadays because of its correlation benefits to any society or nation’s security. It is found that many studies adopted supervised learning approaches to the field of crime prediction compared to others.

It is obviously concluded, that data mining methods achieved the highest crime prediction accuracies, overcoming machine learning methods. Regardless of this, on average, the machine learning out performs data mining in crime prediction. But, when we use the standard deviation of crime prediction accuracies of machine learning and data mining, we can say that the machine learning algorithms perform better than the data mining algorithms.

Eventually, it can be concluded that the comparison of machine learning and data mining algorithms for crime prediction systems give certain indications, such as the selection of an algorithm may depend on the dataset type (like image, text, video, or voice dataset), and there are certain algorithms that preform perfectly on average, but can fail working with other datasets. Crime prediction methods adopting deep learning algorithms were not covered through this survey for time limitation reasons.

Acknowledgements

The authors, want to thank all researchers whose works have been cited in this survey, in the field of crime prediction.

Funding information : This project is funded by the authors only.

Conflict of interest : The authors declare that there is no conflict of interest regarding the publication of this article.

[1] Safat W, Asghar S, Gillani SA. Empirical analysis for crime prediction and forecasting using machine learning and deep learning techniques. IEEE Access J. 2021;9:70080–94. 10.1109/ACCESS.2021.3078117 Search in Google Scholar

[2] Kounadi O, Ristea A, Araujo A, Leitner M. A systematic review on spatial crime forecasting. Crime Sci. 2020;9(1):1–22. 10.1186/s40163-020-00116-7 Search in Google Scholar PubMed PubMed Central

[3] Tollenaar N, van der Heijden PGM. Which method predicts recidivism best?: a comparison of statistical, machine learning and data mining predictive models. J R Stat Soc Ser A. 2013;176(2):565–84. 10.1111/j.1467-985X.2012.01056.x Search in Google Scholar

[4] Enzmann D, Podana Z. Official crime statistics and survey data: Comparing trends of youth violence between 2000 and 2006 in cities of the Czech Republic, Germany, Poland, Russia, and Slovenia. Eur J Crim Policy Res. 2010;16(3):191–205. 10.1007/s10610-010-9121-z Search in Google Scholar

[5] Holst A, Bjurling B. A Bayesian parametric statistical anomaly detection method for finding trends and patterns in criminal behavior. In 2013 European Intelligence and Security Informatics Conference. IEEE; 2013. 10.1109/EISIC.2013.19 Search in Google Scholar

[6] Brunsdon C, Corcoran J, Higgs G. Visualising space and time in crime patterns: A comparison of methods. Comput Environ Urban Syst. 2007;31(1):52–75. 10.1016/j.compenvurbsys.2005.07.009 Search in Google Scholar

[7] Vural MS, Gök M, Yetgin Z. Generating incident-level artificial data using GIS based crime simulation. In 2013 International Conference on Electronics, Computer and Computation (ICECCO). IEEE; 2013. 10.1109/ICECCO.2013.6718273 Search in Google Scholar

[8] Xiang Y, Chau M, Atabakhsh H, Chen H. Visualizing criminal relationships: Comparison of a hyperbolic tree and a hierarchical list. Decis Support Syst. 2005;41(1):69–83. 10.1016/j.dss.2004.02.006 Search in Google Scholar

[9] Jain LC, Seera M, Lim CP, Balasubramaniam P. A review of online learning in supervised neural networks. Neural Comput Appl. 2014;25(3):491–509. 10.1007/s00521-013-1534-4 Search in Google Scholar

[10] EL Aissaoui O, EL Madani EY, Oughdir L, EL Allioui Y. Combining supervised and unsupervised machine learning algorithms to predict the learners’ learning styles. Procedia Comput Sci. 2019;148:87–96. 10.1016/j.procs.2019.01.012 Search in Google Scholar

[11] Mackenzie DM. CDUL: class directed unsupervised learning. Neural Comput Appl. 1995;3(1):2–16. 10.1007/BF01414172 Search in Google Scholar

[12] Rossmo DK, Laverty I, Moore B. Geographic profiling for serial crime investigation, in Geographic information systems and crime analysis. IGI Glob. 2005;6:102–17. 10.4018/978-1-59140-453-8.ch006 Search in Google Scholar

[13] Ristea A, Leitner M. Urban crime mapping and analysis using GIS. ISPRS Int J Geo-Information. 2020;9(9):511. 10.3390/ijgi9090511 Search in Google Scholar

[14] Corcoran JJ, Wilson ID, Ware JA. Predicting the geo-temporal variations of crime and disorder. Int J Forecast. 2003;19(4):623–34. 10.1016/S0169-2070(03)00095-5 Search in Google Scholar

[15] Sangani A, Sampat C, Pinjarkar V. Crime prediction and analysis. In 2nd International Conference on Advances in Science & Technology (ICAST); 2019. 10.2139/ssrn.3367712 Search in Google Scholar

[16] Wang Y, Peng X, Bian J. Computer crime forensics based on improved decision tree algorithm. J Netw. 2014;9(4):1005. 10.4304/jnw.9.4.1005-1011 Search in Google Scholar

[17] Khan M, Ali A, Alharbi Y. Predicting and preventing crime: A crime prediction model using San Francisco crime data by classification techniques. New Jersey: Wiley/Hindawi. Vol. 2022, No. 4830411, 2022. p. 13. 10.1155/2022/4830411 Search in Google Scholar

[18] Ewart BW, Oatley GC. Applying the concept of revictimization: using burglars’ behaviour to predict houses at risk of future victimization. Int J Police Sci Manag. 2003;5(2):69–84. 10.1350/ijps.5.2.69.14324 Search in Google Scholar

[19] Box GEP, Jenkins GM, Reinsel GC, Ljung GM. Time series analysis: forecasting and control. John Wiley & Sons; 2015. Search in Google Scholar

[20] Jangra M, Kalsi S. Naïve Bayes approach for the crime prediction in Data Mining. Int J Comput Appl. 2019;178(4):33–7. 10.5120/ijca2019918907 Search in Google Scholar

[21] Khairuddin A, Alwee R, Haron H. A comparative analysis of artificial intelligence techniques in forecasting violent crime rate. In IOP Conference Series: Materials Science and Engineering. IOP Publishing; 2020. Search in Google Scholar

[22] Sardana D, Marwaha S, Bhatnagar R. Supervised and unsupervised machine learning methodologies for crime pattern analysis. Int J Artif Intell Appl. 2021;12(1):43–58. 10.5121/ijaia.2021.12106 Search in Google Scholar

[23] Sivanagaleela B, Rajesh S. Crime analysis and prediction using fuzzy c-means algorithm. In 3rd International Conference on Trends in Electronics and Informatics (ICOEI). IEEE; 2019. 10.1109/ICOEI.2019.8862691 Search in Google Scholar

[24] Liu X, Sun H, Han S, Han S, Niu S, Qin W, et al. A data mining research on office building energy pattern based on time-series energy consumption data. Energy Build. 2022;259:111888. 10.1016/j.enbuild.2022.111888 Search in Google Scholar

[25] Borowik G, Wawrzyniak ZM, Cichosz P. Time series analysis for crime forecasting. In 2018 26th International Conference on Systems Engineering (ICSEng). IEEE; 2018. 10.1109/ICSENG.2018.8638179 Search in Google Scholar

[26] Goel A, Singh B. White collar crimes: A study in the context of classification, causation and preventive measures. Contemp Soc Sci. 2018;27:84–92. 10.29070/27/58311 Search in Google Scholar

[27] Grove L, Farrell G. Once bitten, twice shy: Repeat victimization and its prevention. Oxf Handb Crime Prev. 2012;404–19. 10.1093/oxfordhb/9780195398823.013.0020 Search in Google Scholar

[28] Chainey SP, da Silva BFA. Examining the extent of repeat and near repeat victimisation of domestic burglaries in Belo Horizonte, Brazil. Crime Sci. 2016;5(1):1–10. 10.1186/s40163-016-0049-6 Search in Google Scholar

[29] Peter P, Ickjai L. Crime analysis through spatial areal aggregated density patterns. Geoinformatica. 2011;15(1):49–74. 10.1007/s10707-010-0116-1 Search in Google Scholar

[30] Jonas p, Paul E, Stijn V, Marc MVH, Guido D. Gaining insight in domestic violence with emergent self organizing maps. Expert Syst Appl. 2009;36(9):11864–74. 10.1016/j.eswa.2009.04.027 Search in Google Scholar

[31] Kang HW, Kang HB. Prediction of crime occurrence from multi-modal data using deep learning. PLoS One. 2017;12(4):e0176244. 10.1371/journal.pone.0176244 Search in Google Scholar PubMed PubMed Central

[32] Shekhar S, Evans MR, Kang JM, Mohan P. Identifying patterns in spatial information: A survey of methods. Wiley Interdiscip Reviews Data Min Knowl Discovery. 2011;1(3):193–214. 10.1002/widm.25 Search in Google Scholar

[33] Mookiah L, Eberle W, Siraj A. Survey of crime analysis and prediction. In The Twenty-Eighth International Flairs Conference; 2015. Search in Google Scholar

[34] Hassani H, Huang X, Silva ES, Ghodsi M. A review of data mining applications in crime. Stat Anal Data Mining: ASA Data Sci J. 2016;9(3):139–54. 10.1002/sam.11312 Search in Google Scholar

[35] Falade A, Azeta A, Oni A, Odun-ayo I. Systematic literature review of crime prediction and data mining. Rev Comput Eng Stud. 2019;6(3):56–63. 10.18280/rces.060302 Search in Google Scholar

[36] Okeke OC. An overview of crime analysis, prevention and predicton using data mining based on real time and location data. Int J Recent Technol Eng. 2022;5(10):99–103. 10.33564/IJEAST.2021.v05i10.015 Search in Google Scholar

[37] Kianmehr K, Alhajj R. Crime hot-spots prediction using support vector machine. In IEEE International Conference on Computer Systems and Applications. IEEE Computer Society; 2006. 10.1109/AICCSA.2006.205203 Search in Google Scholar

[38] Antolos D, Liu D, Ludu A, Vincenzi D. Burglary crime analysis using logistic regression. In International Conference on Human Interface and the Management of Information. Berlin: Springer; 2013. 10.1007/978-3-642-39226-9_60 Search in Google Scholar

[39] Cavadas B, Branco P, Pereira S. Crime prediction using regression and resources optimization. In Portuguese Conference on Artificial Intelligence. Springer; 2015. 10.1007/978-3-319-23485-4_51 Search in Google Scholar

[40] Cesario E, Catlett C, Talia D. Forecasting crimes using autoregressive models. in 2016 IEEE 14th Intl Conf on Dependable, Autonomic and Secure Computing, 14th Intl Conf on Pervasive Intelligence and Computing, 2nd Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology Congress (DASC/PiCom/DataCom/CyberSciTech). IEEE; 2016. 10.1109/DASC-PICom-DataCom-CyberSciTec.2016.138 Search in Google Scholar

[41] Vural MS, Gök M. Criminal prediction using Naive Bayes theory. Neural Comput Appl. 2017;28(9):2581–92. 10.1007/s00521-016-2205-z Search in Google Scholar

[42] Hou M, Hu X, Cai J, Han X, Yuan S. An integrated graph model for spatial–temporal urban crime prediction based on attention mechanism. ISPRS Int J Geo-Information. 2022;11(5):294. 10.3390/ijgi11050294 Search in Google Scholar

[43] Ilhan F, Tekin SF, Aksoy B. Spatio-temporal crime prediction with temporally hierarchical convolutional neural networks. In 2020 28th Signal Processing and Communications Applications Conference (SIU). IEEE; 2020. 10.1109/SIU49456.2020.9302169 Search in Google Scholar

[44] Meskela TE, Afework YK, Ayele NA, Teferi MW, Mengist TB. Designing time series crime prediction model using long short-term memory recurrent neural network. Int J Recent Technol Eng. 2020;9:402–5. 10.35940/ijrte.D5025.119420 Search in Google Scholar

[45] Hussain FS, Aljuboori AF. A crime data analysis of prediction based on classification approaches. Baghdad Sci J. 2022;4:1073–7. Search in Google Scholar

[46] Lin YL, Yen MF, Yu LC. Grid-based crime prediction using geographical features. ISPRS Int J Geo-Information. 2018;7(8):298. 10.3390/ijgi7080298 Search in Google Scholar

[47] Stec A, Klabjan D. Forecasting crime with deep learning. arXiv preprint arXiv; 2018. p. 01486. Search in Google Scholar

[48] Kim KS, Jeong YH. A study on crime prediction to reduce crime rate based on artificial intelligence. Korea J Artif Intell. 2021;9(1):15–20. Search in Google Scholar

[49] Bogomolov A, Lepri B, Staiano J, Oliver N, Pianesi F, Pentland A. Once upon a crime: towards crime prediction from demographics and mobile data. In Proceedings of the 16th International Conference on Multimodal Interaction; 2014. 10.1145/2663204.2663254 Search in Google Scholar

[50] Zhuang Y, Almeida M, Morabito M Ding W. Crime hot spot forecasting: A recurrent model with spatial and temporal information. In 2017 IEEE International Conference on Big Knowledge (ICBK). IEEE; 2017. 10.1109/ICBK.2017.3 Search in Google Scholar

[51] Ivan N, Ahishakiye E, Omulo EO, Taremwa D. Crime prediction using decision tree (J48) classification algorithm. International Journal of Computer and Information Technology. 2017;6:188–95. Search in Google Scholar

[52] El Bour HA, Ounacer S, Elghomari Y, Jihal H, Azzouazi M. A crime prediction model based on spatial and temporal data. Periodicals Eng Nat Sci. 2018;6(2):360–4. 10.21533/pen.v6i2.524 Search in Google Scholar

[53] Kim S, Joshi P, Kalsi PS, Taheri P. Crime analysis through machine learning. In IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON). IEEE; 2018. 10.1109/IEMCON.2018.8614828 Search in Google Scholar

[54] Bharati A, RA KS. Crime prediction and analysis using machine learning. Int Res J Eng Technol (IRJET). 2018;5:1037–42. Search in Google Scholar

[55] Mahmud S, Nuha M, Sattar A. Crime Rate Prediction Using Machine Learning and Data Mining, in Soft Computing Techniques and Applications. Singapore: Springer; 2021. p. 59–69. 10.1007/978-981-15-7394-1_5 Search in Google Scholar

[56] Almuhanna AA, Alrehili MM, Alsubhi SH, Syed L. Prediction of crime in neighbourhoods of New York City using spatial data analysis. In 2021 1st International Conference on Artificial Intelligence and Data Analytics (CAIDA). IEEE; 2021. 10.1109/CAIDA51941.2021.9425120 Search in Google Scholar

[57] Yu CH, Ward MW, Morabito M, Ding W. Crime forecasting using data mining techniques. In 2011 IEEE 11th International Conference on Data Mining Workshops. IEEE; 2011. 10.1109/ICDMW.2011.56 Search in Google Scholar

[58] Sharma H, Kumar S. A survey on decision tree algorithms of classification in data mining. Int J Sci Res. 2016;5(4):2094–7. 10.21275/v5i4.NOV162954 Search in Google Scholar

[59] Gupta A, Mohammad A, Syed A, Halgamuge MN. A comparative study of classification algorithms using data mining: crime and accidents in Denver City the USA. Education. 2016;7(7):374–81. 10.14569/IJACSA.2016.070753 Search in Google Scholar

[60] Boppuru PR, Ramesha K. Spatio-temporal crime analysis using KDE and ARIMA models in the indian context. Int J Digital Crime Forensics. 2020;12(4):1–19. 10.4018/IJDCF.2020100101 Search in Google Scholar

[61] Tayal D, Jain A, Arora S, Agarwal S, Gupta T, Tyagi N. Crime detection and criminal identification in India using data mining techniques. AI Soc. 2015;30(1):117–27. 10.1007/s00146-014-0539-6 Search in Google Scholar

[62] Iqbal R, Murad MAA, Mustapha A, Panahy PHS, Khanahmadliravi N. An experimental study of classification algorithms for crime prediction. Indian J Sci Technol. 2013;6(3):4219–25. 10.17485/ijst/2013/v6i3.6 Search in Google Scholar

[63] Almanie T, Mirza R, Lor E. Crime prediction based on crime types and using spatial and temporal criminal hotspots. arXiv preprint arXiv; 2015. p. 02050. 10.5121/ijdkp.2015.5401 Search in Google Scholar

[64] Yerpude P, Gudur V. Predictive modelling of crime dataset using data mining. Int J Data Min Knowl Manag Process. 2020;7:83–99. Search in Google Scholar

[65] Prathap BR, Krishna A, Balachandran K. Crime analysis and forecasting on spatio temporal news feed data—An indian context, in artificial intelligence and blockchain for future cybersecurity applications. Switzerland: Springer; 2021. p. 307–27. 10.1007/978-3-030-74575-2_16 Search in Google Scholar

© 2023 the author(s), published by De Gruyter

This work is licensed under the Creative Commons Attribution 4.0 International License.

• X / Twitter

Supplementary Materials

Please login or register with De Gruyter to order this product.

Journal and Issue

Articles in the same issue.

• DOI: 10.1109/ACCESS.2023.3286344
• Corpus ID: 257805247

Crime Prediction Using Machine Learning and Deep Learning: A Systematic Review and Future Directions

• V. Mandalapu , Lavanya Elluri , +1 author Nirmalya Roy
• Published in IEEE Access 28 March 2023
• Computer Science, Law

Figures and Tables from this paper

23 Citations

Crime prediction by comparing machine learning and deep learning algorithms, deep learning based crime prediction models: experiments and analysis, a study on smart machine learning (ml) tools for crime detection and prediction, evaluating machine learning models best fit for crime prediction in windhoek, a repository based criminal identification - evolution in forensics using deep learning, analysis and prediction of crimes against women, context-based crime detection : a framework integrating computer vision technologies, research on victim genders in la based-on machine learning, a systematic review of using machine learning and natural language processing in smart policing, algorithmic guardians: evaluating machine learning for predicting criminal activities, 82 references, empirical analysis for crime prediction and forecasting using machine learning and deep learning techniques, crime prediction model using deep neural networks, developing machine learning based predictive models for smart policing, an empirical analysis of machine learning algorithms for crime prediction using stacked generalization: an ensemble approach, crime prediction using spatio-temporal data, a novel multi-module approach to predict crime based on multivariate spatio-temporal data using attention and sequential fusion model, crime forecasting: a machine learning and computer vision approach to crime prediction and prevention, crime prediction & monitoring framework based on spatial analysis, crime data analysis and prediction of perpetrator identity using machine learning approach, crime prediction and monitoring in porto, portugal, using machine learning, spatial and text analytics, related papers.

Showing 1 through 3 of 0 Related Papers

Advertisement

Crime analysis and prediction using machine-learning approach in the case of Hossana Police Commission

• Original Article
• Published: 01 March 2024

Cite this article

• Betelhem Zewdu Wubineh   ORCID: orcid.org/0000-0002-4790-7449 1 , 2  

168 Accesses

Explore all metrics

Crime is a socioeconomic problem that affects the quality of life and economic growth of a country, and it continues to increase. Crime prevention and prediction are systematic approaches used to locate and analyze historical data to identify trends that can be employed in identifying crimes and criminals. The objective of this study is to predict the type of crime that occurred in the city and identify the important features that make this prediction using a machine learning technique. For this experimental investigation, a supervised learning method was used to classify the types of crimes based on the final labelled class that indicates which type of crime is committed. Thus, decision tree (DT), random forest (RF), and K-nearest neighbor (KNN) algorithms are utilized along with the Python programming language in the Jupyter notebook environment. A total of 1400 records and nine attributes were used, and the data were split into training and testing sets, with 80% allocated to training and 20% for testing. The decision tree achieved an accuracy score of 84%, followed by the random forest at 86.07% and K-nearest neighbor at 81%. Besides this, the job of the offender, the victim’s age, and the offender’s age are the important features that cause crime. Therefore, it can be concluded that the use of machine learning to analyze historical data and the random forest algorithm to classify crimes yields promising results in predicting the type of crime.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Crime Prediction Using Data Mining and Machine Learning

Crime Prediction Using Spatio-Temporal Data

A Classification Approach for Crime Prediction

Explore related subjects.

• Artificial Intelligence

Alasadi, Suad A., and Wesam S. Bhaya. 2017. Review of Data Preprocessing Techniques.Pdf. Journal of Engineering and Applie Sciences .

Barnadas, Miquel Vaquero. 2016. Escola Tècnica d ’ Enginyeria de Telecomunicació de Barcelona Advisors : Margarita Cabrera Bean and José Adrian Rodríguez, no. September.

Chomboon, Kittipong, Pasapitch Chujai, Pongsakorn Teerarassamee, Kittisak Kerdprasop, and Nittaya Kerdprasop. 2015. An Empirical Study of Distance Metrics for K-Nearest Neighbor Algorithm. In Proceedings of the 3rd International Conference on Industrial Application Engineering , 280–85.

Dalianis, Hercules. 2018. Evaluation Metrics and Evaluation. Clinical Text Mining, 1967: 45–53. https://doi.org/10.1007/978-3-319-78503-5_6 .

Article   Google Scholar  

Ghosh, Sourav Kumar, and Farhatul Janan. 2021. Prediction of Student’s Performance Using Random Forest Classifier. Proceedings of the International Conference on Industrial Engineering and Operations Management , 7089–7100.

Hassani, Hossein, Xu Huang, Emmanuel S. Silva, and Mansi Ghodsi. 2016. A Review of Data Mining Applications in Crime . Statistical Analysis and Data Mining , vol. 9. https://doi.org/10.1002/sam.11312 .

Khan, Muzammil, Azmat Ali, and Yasser Alharbi. 2022. Predicting and Preventing Crime: A Crime Prediction Model Using San Francisco Crime Data by Classification Techniques. Complexity . https://doi.org/10.1155/2022/4830411 .

Kim, Suhong, Param Joshi, Parminder Singh Kalsi, and Pooya Taheri. 2019. Crime Analysis Through Machine Learning. 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference, IEMCON 2018 , no. January: 415–20. https://doi.org/10.1109/IEMCON.2018.8614828 .

Kuhkan, Maryam. 2016. A Method to Improve the Accuracy of K-Nearest Neighbor Algorithm. International Journal of Computer Engineering and Information Technology 8 (6): 90–95. www.ijceit.org .

Lin, Ying Lung, Meng Feng Yen, and Yu. Liang Chih. 2018. Grid-Based Crime Prediction Using Geographical Features. ISPRS International Journal of Geo-Information . https://doi.org/10.3390/ijgi7080298 .

Liu, Jialiang. 2022. Crime Prediction from Digital Videos Using Deep Learning. Auckland University of Technology.

Llaha, Olta. 2020. Crime Analysis and Prediction Using Machine Learning. 2020 43rd International Convention on Information, Communication and Electronic Technology, MIPRO 2020 - Proceedings , no. September 2020: 496–501. https://doi.org/10.23919/MIPRO48935.2020.9245120 .

McClendon, Lawrence, and Natarajan Meghanathan. 2015. Using Machine Learning Algorithms to Analyze Crime Data. Machine Learning and Applications: An International Journal 2 (1): 1–12. https://doi.org/10.5121/mlaij.2015.2101 .

Nair, Swati, Saloni Soniminde, Sruthi Sureshbabu, Apurva Tamhankar, and Sagar Kulkarni. 2019. Assist Crime Prevention Using Machine Learning. SSRN Electronic Journal . https://doi.org/10.2139/ssrn.3349683 .

Ogunde, Adewale Opeoluwa, Gabriel Opeyemi Ogunleye, and Oluwaleke Oreoluwa. 2017. A Decision Tree Algorithm Based System for Predicting Crime in the University. Machine Learning Research 2 (1): 26–34. https://doi.org/10.11648/j.mlr.20170201.14 .

Powers, David M. W. 2020. Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness and Correlation, no. January 2011. https://doi.org/10.9735/2229-3981 .

Rahmawati, Sinta Denovi, and Yova Ruldeviyani. 2019. Data Quality Management Strategy to Improve the Quality of Worker’s Wage and Income Data: A Case Study in BPS-Statistics Indonesia, 2018. In 2019 Fourth International Conference on Informatics and Computing (ICIC) , 1–6. IEEE.

Saltos, Ginger, and Mihaela Cocea. 2017. An Exploration of Crime Prediction Using Data Mining on Open Data. International Journal of Information Technology and Decision Making 16 (5): 1155–1181. https://doi.org/10.1142/S0219622017500250 .

Schlegel, Udo. 2018. Towards Crime Forecasting Using Deep Learning. Universität Konstanz Department of Computer Science Master Thesis .

Song, Yan-Yan., and L.U. Ying. 2015. Decision Tree Methods: Applications for Classification and Prediction. Shanghai Archives of Psychiatry 27 (2): 130.

PubMed   PubMed Central   Google Scholar  

Stalidis, Panagiotis, Theodoros Semertzidis, and Petros Daras. 2021. Examining Deep Learning Architectures for Crime Classification and Prediction. Forecasting 3 (4): 741–762. https://doi.org/10.3390/forecast3040046 .

Thongsatapornwatana, Ubon. 2016. A Survey of Data Mining Techniques for Analyzing Crime Patterns. 2016 2nd Asian Conference on Defence Technology, ACDT 2016 , no. January: 123–28. https://doi.org/10.1109/ACDT.2016.7437655 .

Tolan, Ghada M., and Omar S. Soliman. 2015. An Experimental Study of Classification Algorithms for Terrorism Prediction. International Journal of Knowledge Engineering-IACSIT 1 (2): 107–112. https://doi.org/10.7763/ijke.2015.v1.18 .

Walczak, Steven. 2021. Predicting Crime and Other Uses of Neural Networks in Police Decision Making. Frontiers in Psychology . https://doi.org/10.3389/fpsyg.2021.587943 .

Article   PubMed   PubMed Central   Google Scholar  

Yerpude, Prajakta, and Vaishnavi Gudur. 2017. Predictive Modelling of Crime Dataset Using Data Mining. International Journal of Data Mining & Knowledge Management Process 7 (4): 43–58. https://doi.org/10.5121/ijdkp.2017.7404 .

Zahran, Samah, Eman Mohamed, and Hamdy Mousa. 2021. Detecting and Predicting Crimes Using Data Mining Techniques: Comparative Study. IJCI. International Journal of Computers and Information 8 (2): 57–62. https://doi.org/10.21608/ijci.2021.207749 .

Download references

Author information

Authors and affiliations.

Faculty of Information and Communication Technology, Wroclaw University of Science and Technology, Wrocław, Poland

Betelhem Zewdu Wubineh

Department of Information Technology, College of Engineering and Technology, Wachemo University, Hosaena, Ethiopia

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Betelhem Zewdu Wubineh .

Ethics declarations

Conflict of interest.

The authors declare there are no financial and competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Wubineh, B.Z. Crime analysis and prediction using machine-learning approach in the case of Hossana Police Commission. Secur J (2024). https://doi.org/10.1057/s41284-024-00416-6

Download citation

Accepted : 06 January 2024

Published : 01 March 2024

DOI : https://doi.org/10.1057/s41284-024-00416-6

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Machine learning
• Random forest
• Find a journal
• Publish with us
• Track your research