Abstract
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.
Introduction
In the realm of public safety, the paradigm is shifting from reactive enforcement to proactive prevention. The central challenge lies in forecasting criminal activity—a "wicked problem" characterized by high-dimensional data, complex human behaviors, and significant ethical implications. While modern data sources such as surveillance feeds, social media, and demographic records are more accessible than ever, the research landscape remains fragmented. This systematic review aims to bridge the gap between theoretical AI research and practical tools for law enforcement by identifying the most effective algorithms and available datasets.
Systematic Review Methodology
To ensure a rigorous analysis, the authors conducted a multi-stage systematic literature review. The process began with an automated initial search across metadata attributes, yielding nearly 10,000 results. Through successive rounds of filtering—including full-text searches, index term analysis, and manual reading—the pool was narrowed down to 51 high-impact core studies that form the basis of this analysis.
Data Sources and Algorithmic Approaches
The research identifies a diverse array of data sources utilized in crime forecasting. These range from city-specific open data portals (e.g., Chicago, NYC, London) and national repositories to non-traditional sources like sentiment data from Twitter and geospatial points-of-interest from Foursquare. Multimedia data, including CCTV footage, is increasingly used for real-time weapon and violence detection.
Methodologically, the field is divided into two primary categories:
- Machine Learning (ML): Traditional models like Random Forest, SVM, and XGBoost remain dominant, accounting for 67% of the surveyed studies. Their popularity is driven by their performance on smaller datasets and their relative interpretability.
- Deep Learning (DL): More complex models like CNNs (for vision tasks) and LSTMs (for temporal dependencies) are rising. Hybrid models, such as Spatio-Temporal Graph Convolutional Networks (ST-GCN), are becoming the gold standard for capturing the intersection of geography and time.
Emerging Trends and Performance Metrics
A notable finding of this review is the "COVID effect" on research output. There was a significant spike in publications during 2020 and 2021, likely due to researchers exploring how neighborhood crime dynamics shifted during global lockdowns.
In terms of performance, the results are promising:
- Classification Dominance: 63% of studies focus on classification (e.g., identifying high-risk zones). XGBoost models have achieved up to 89% accuracy in specific regional studies.
- Regression and Vision: Random Forest Regressors reached 97% accuracy in predicting homicide impacts, while hybrid CNN-GRU models for real-time weapon detection demonstrated a 95.97% accuracy rate.
Challenges and Future Outlook
Despite technical successes, the "black box" nature of advanced AI remains a significant hurdle. In law enforcement, transparency is vital; an algorithm must not only predict a hotspot but also explain the underlying factors. The review highlights Explainable AI (XAI) tools like SHAP as a critical area for future development.
Furthermore, the field must address data scarcity in certain regions and the inherent biases in historical crime data. Future research directions include the exploration of Reinforcement Learning, Transfer Learning, and Generative Adversarial Networks (GANs), alongside a rigorous focus on the ethical implications of predictive policing to ensure these technologies do not perpetuate systemic discrimination.
Summary updated