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

• Machine learning methods analyze crime data to predict patterns and relate incidents to environmental and demographic factors such as location, weather, and time of day.Decision trees, random forests, and support vector machines are examples of methods trained on city crime data.Source: §1 · p. 2
• Predictive policing uses geographic crime data to identify hotspots and predict incidents, informing where law-enforcement resources are directed.The review describes this as an application of machine learning.Source: §1 · p. 2
• Deep learning approaches use spatial or temporal crime data, while computer vision and video analysis can detect and classify activities such as vandalism, theft, and assault.The review also describes combining crime reports, social media, and police records for a broader view of criminal activity.Source: §1 · pp. 2–3
• Crime prediction faces challenges of data availability and reliability, privacy and ethics, and model interpretability.The paper calls for interpretable models that explain their predictions.Source: §1 · p. 3
• The review amalgamates machine learning and deep learning studies, identifies publicly available neighborhood-crime datasets, and proposes future research directions.These contributions aim to extend the literature base and archive data resources for future scholars.Source: §1 · p. 3

2 Research Methodology

• The review searches multiple research databases for studies on crime prediction and uses both automated searching and manual reading to select relevant work.The stated aim includes identifying datasets used to apply algorithms.Source: §2 · p. 4
• The search queries used crime-related terms and alternatives for prediction, detection, recognition, machine learning, deep learning, clustering, and natural language processing.Queries were adapted to the syntax of IEEE, Science Direct, and ACM databases.Source: §2 · p. 4

2.1 Data Collection

• The search applies filters to database results, reducing a pool of more than 450 papers to 157 after document-title filtering.The review searched individual research libraries rather than Google Scholar to avoid duplicates.Source: §2.1 · pp. 4–5
• After automated filtering, the authors manually read papers and assessed their focus, datasets, algorithms, and accuracy rates to select the final set.The review also selected 30 papers for the main text and 20 for appendix A1.Source: §2 · p. 4, §2.1 · p. 6

3.1 Pre-assessment Literature Analysis

• IEEE publication trends rose from 2018 to 2021 and fell in 2022; the authors suggest COVID-19’s 2020–2021 period could help explain the rise.Source: Figure 3 · p. 6, §3.1 · p. 6
• Five pages was the most common article length, with 46 articles; very few exceeded ten pages.Source: Figure 4 · p. 7, §3.1 · p. 7
• Conference papers received 60 citations, compared with 10 for journal articles, 1 for book chapters, 3 for symposiums, and none for seminars or workshops.Source: Figure 5 · p. 7, §3.1 · p. 7
• Conferences accounted for 82% of publications, followed by journals at 9%, symposiums at 4%, book chapters at 3%, and seminars and workshops at 1% each.Source: Figure 6 · p. 8, §3.1 · p. 8

3.2 Post-assessment Literature Analysis

• The word cloud highlights crime-detection themes such as “crime,” “criminal,” “policing,” “enforcement,” and “security,” alongside “prediction,” “algorithms,” and “techniques.”Source: §3.2 · pp. 8–9, Figure 7 · p. 9
• 67% of articles used ML, 21% DL, 8% ML+DL, 2% DL+NLP, and 2% ML+NLP, making ML the most-used technique type.Source: Figure 8 · p. 9, §3.2 · pp. 9–10
• 63% of articles focused on classification, 29% on regression, 6% on clustering, and 2% on combined classification and clustering.Source: §3.2 · p. 10
• All five technique types were used for classification; ML alone was used for clustering and combined clustering-classification, while NLP also appeared in classification and regression.Source: §3.2 · p. 10, Figure 10 · p. 11

4 Crime Prediction Process & Datasets

• The pipeline moves from collecting crime, demographic, and weather data through preprocessing, train/test splitting, feature engineering, model training, and performance evaluation.The evaluated results can support law-enforcement and crime-prevention decisions.Source: §4 · p. 11, Figure 11 · p. 11
• Datasets include Chicago, London, Los Angeles, New York City, and Philadelphia crime records, alongside global CCTV, aggression, and weapons datasets for real-time prediction research.Source: §4 · p. 13
• Table 1 lists locations and access information for surveyed datasets, which researchers can use to build models intended to support crime prevention and response.Source: Table 1 · p. 12, §4 · p. 13

5 Crime Prediction using Machine Learning techniques

• Traditional machine-learning approaches include decision trees, support vector machines, logistic regression, and random forests, which the review describes as requiring fewer data and being easier to interpret than deep learning.Source: §5 · p. 14

5.0.1 Machine Learning based Regression Methods for Crime Prediction

• Regression studies apply ARIMAX and hotspot forecasting, compare urban crime-rate models, and report random-forest analysis of city indicators including unemployment and illiteracy.Source: §5.0.1 · pp. 14–15

5.0.2 Machine Learning based Classification Methods for Crime Prediction

• Classification studies span incremental crime-report learning, XGBoost and SHAP feature analysis, psychiatric risk screening, and graph ensembles reported to outperform traditional classifiers.One psychiatric-care study identified 47% of future forensic psychiatric patients and made correct predictions in 57% of samples.Source: §5.0.2 · pp. 16–19

5.1 Crime Prediction using Deep Learning techniques

• Deep-learning studies use CNNs, deep neural networks, and sentiment analysis across text, images, audio, and social media for crime-related regression and classification tasks.Source: §5.1 · p. 19

5.1.1 Deep Learning based Regression Methods for Crime Prediction

• Across five reviewed papers, regression is integrated with neural and fusion methods to combine crime data with external variables; LSTM and ST-GCN capture temporal and spatial dependencies, respectively.Source: §5.1.1 · pp. 20, 22

5.1.2 Deep Learning based Classification Methods for Crime Prediction

• Deep-learning classification spans images, text, audio, and social media, including CNN/RNN studies of scene objects and behavior, anomaly detection, and audio-text event classification.Source: §5.1.2 · pp. 22, 24

6 Discussion and future work

• These methods can analyze large datasets to identify patterns in criminal activity, including signals from social media and other online sources.The discussion says such data can provide insights into criminal activities that are yet to be committed.Source: §6 · p. 24
• CNNs and RNNs have been used to analyze security-camera footage, while real-time models may help agencies respond quickly to incidents.The discussion also identifies wearable devices such as body cameras and smartwatches as potential sources of crime-related data.Source: §6 · pp. 24–25
• Crime prediction needs interpretable models and causal explanations, since incorrect predictions may have serious consequences for individuals and communities.The discussion distinguishes model-based explanations from explanations focused on cause-and-effect relationships.Source: §6 · p. 25
• Researchers need accurate, up-to-date data and algorithms that integrate multiple sources, including demographic data outside crime-specific datasets.Source: §6 · p. 25
• Further work is needed on bias, discrimination, privacy, real-world effectiveness, and the scalability of large-scale implementations.The discussion calls for more rigorous evaluations of accuracy and effectiveness in real-world scenarios.Source: §6 · p. 25
• The authors conclude that addressing these gaps could advance understanding of crime-prediction methods and contribute to more effective and efficient policing strategies.Source: §6 · pp. 25–26
• The review identifies eight technique- and feature-oriented research questions, including reinforcement learning and visual features for neighborhood-crime detection.It notes that relevant datasets are available for pursuing early recognition of neighborhood crime.Source: §6 · p. 26

7 Conclusion

• The paper examines more than 150 articles and studies 51 selected articles to identify machine learning and deep learning techniques and publicly available datasets.Source: §7 · p. 27
• The reviewed approaches show promise for crime prediction, but the authors note a lack of literature on how these technologies can be applied to the problem.The conclusion associates this promise with the use of large datasets and sophisticated algorithms.Source: §7 · p. 27
• The authors present their findings, datasets, and future directions as resources for the existing crime-prediction research community.Source: §7 · p. 27

Appendix A Additional papers on crime detection using Machine Learning and Deep Learning

• The appendix lists studies using spatial analysis, clustering, association-rule mining, classification, and time-series methods.Source: Table A1 · p. 28, Appendix A · pp. 29–30
• One study reports that support vector regression significantly outperformed polynomial and auto-regressive approaches.The listed approaches are polynomial, auto-regressive, and support vector regression.Source: Appendix A · p. 29
• Reported outcomes include RMSE of 0–10 for most crimes and mean absolute error (MAE) of 11.5.Source: Appendix A · p. 29
• Other listed results include accuracy of 70 to 80% and a true positive rate of 89.17%.Source: Appendix A · pp. 30–31
• The listed datasets span settings including NYC and Chicago crime data and 97,287 Australian Federal Police drug-case images.Source: Appendix A · pp. 30–31