Scholarlink

<p>Urban crime prediction requires robust, detailed structures capable of capturing both spatial and temporal dynamics while remaining fair and interpreted. This paper introduces a Hierarchical Spatiotemporal Transformer (HSTT) that uses dynamic graph learning to enhance crime risk estimates for practical policing. Geospatial tiling and calendar encoding combined multi-source data such as structured signals, mobility traces, and external factors to generate comprehensive spatiotemporal feature representations. A comparison was made between traditional machine learning methods (like gradient-boosted trees and generalized additive models) and deep learning approaches (including seq2seq transformers and graph neural networks). Results show that HSTT achieved high accuracy with a score of 0.91, an F1-score of 0.90, an AUC of 0.92, and the lowest RMSE at 0.18. Beyond predictive power, HSTT also delivered the best coverage (0.93), the least sharpness (0.15), and the smallest demographic gap (D = 0.04), balancing accuracy and fairness. Although demanding in computational resources, offline training and interpretable results make the framework feasible for real-world application. This study provides an objective perspective on making informed decisions about safer, more transparent city governance by integrating the latest spatiotemporal modeling with ethical considerations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spatiotemporal Crime Prediction and Forecasting Using Machine Learning and Deep Learning Models

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Riyazuddin,
  • Syed Ziaur Rahman,
  • Karnam Sreenu,
  • Mohd Sirajuddin,
  • Syed Mazharuddin,
  • Jaffar Sadiq

摘要

Urban crime prediction requires robust, detailed structures capable of capturing both spatial and temporal dynamics while remaining fair and interpreted. This paper introduces a Hierarchical Spatiotemporal Transformer (HSTT) that uses dynamic graph learning to enhance crime risk estimates for practical policing. Geospatial tiling and calendar encoding combined multi-source data such as structured signals, mobility traces, and external factors to generate comprehensive spatiotemporal feature representations. A comparison was made between traditional machine learning methods (like gradient-boosted trees and generalized additive models) and deep learning approaches (including seq2seq transformers and graph neural networks). Results show that HSTT achieved high accuracy with a score of 0.91, an F1-score of 0.90, an AUC of 0.92, and the lowest RMSE at 0.18. Beyond predictive power, HSTT also delivered the best coverage (0.93), the least sharpness (0.15), and the smallest demographic gap (D = 0.04), balancing accuracy and fairness. Although demanding in computational resources, offline training and interpretable results make the framework feasible for real-world application. This study provides an objective perspective on making informed decisions about safer, more transparent city governance by integrating the latest spatiotemporal modeling with ethical considerations.