Spatio-Temporal-Network Point Processes for Modeling Crime Events with Landmarks

Zheng Dong, Jorge Mateu, Yao XiearXiv 2024 · DOI · arXiv

Abstract: Self-exciting point processes are widely used to model the contagious effects of crime events living within continuous geographic space, using their occurrence time and locations. However, in urban environments, most events are naturally constrained within the city's street network structure, and the contagious effects of crime are governed by such a network geography. Meanwhile, the complex distribution of urban infrastructures also plays an important role in shaping crime patterns across space. We introduce a novel spatio-temporal-network point process framework for crime modeling that integrates these urban environmental characteristics by incorporating self-attention graph neural networks. Our framework incorporates the street network structure as the underlying event space, where crime events can occur at random locations on the network edges. To realistically capture criminal movement patterns, distances between events are measured using street network distances. We then propose a new mark for a crime event by concatenating the event's crime category with the type of its nearby landmark, aiming to capture how the urban design influences the mixing structures of various crime types. A graph attention network architecture is adopted to learn the existence of mark-to-mark interactions. Extensive experiments on crime data from Valencia, Spain, demonstrate the effectiveness of our framework in understanding the crime landscape and forecasting crime risks across regions.

Introduction

Traditional predictive policing models often treat the city as a continuous, two-dimensional plane where distances are measured "as the crow flies" (Euclidean distance). However, urban crime is physically constrained by the city's layout; criminals move along streets, and their escape routes are dictated by the network of roads. Furthermore, most models ignore the context of the crime's location. An assault occurring near a nightclub at 2:00 AM has a very different "contagion" profile than one occurring near a bank during business hours. This research aims to bridge this gap by integrating the physical street network and the functional characteristics of urban landmarks into a unified statistical framework.

The STNPP Framework

The researchers developed the Spatio-Temporal-Network Point Process (STNPP), which builds on the Hawkes process—a mathematical model used to describe events that "self-excite," where one event increases the probability of future events.

Conceptual architecture of the Spatio-Temporal-Network Point Process (STNPP) framework.

Key technical innovations include:

  • Street-Network Distance ($d_{net}$): Instead of straight-line distance, the model calculates the shortest path between events along the actual city street graph ($G$).
  • Graph Attention Networks (GAT): The model uses a GNN architecture to learn the "influence kernel"—the likelihood that a crime of type A in zone X will trigger a crime of type B in zone Y. The GAT uses a self-attention mechanism to identify which combinations of crime and environment are most strongly linked.
  • Stochastic Gradient Descent (SGD) Estimation: To handle the massive computational load of calculating network distances for tens of thousands of events, the authors optimized the model using time-windowed subsequences and SGD.

Urban Context and Event Marks

A central innovation of STNPP is the use of Novel Event Marks. Each crime is assigned a "mark" that concatenates its category (e.g., Assault) with its Urban Functional Zone. These zones are determined by the most common nearby landmarks (Financial, Industrial, Market, Nightclub, Police, Restaurant, or Taxi) using a k-nearest neighbors approach.

Visualization of urban functional zones and the concatenation of crime categories with landmark types.

By combining crime type with landmark type (creating 21 unique "crime-landmark labels"), the model discovered distinct community structures. For example, crimes in industrial and nightclub zones clustered together, reflecting specific patterns of human activity. The study also identified that certain event types are disproportionately influential; for instance, assaults in taxi zones were found to have the strongest triggering effect on subsequent crimes.

Experimental Evaluation

The framework was evaluated using a comprehensive real-world dataset from Valencia, Spain, spanning 2015–2019. This included 47,125 recorded incidents (Assault, Subtraction, and Others), 1,975 city landmarks, and a street network consisting of 8,043 nodes and 12,309 edges.

STNPP was compared against several baselines, including Vector Autoregression (VAR) and the standard Epidemic-Type Aftershock Sequence (ETAS) model:

  • Predictive Accuracy: STNPP achieved the lowest Mean Absolute Error (MAE) in both in-sample estimations and out-of-sample predictions.
  • Next-Event Prediction: In predicting the time, location, and type of the very next crime, STNPP outperformed all baselines, particularly in spatial accuracy (3.231 km MAE vs. 3.649 km for ETAS).
  • Model Fit: The model showed significantly better log-likelihood and lower Akaike Information Criterion (AIC) scores, indicating a superior fit to the complex dynamics of urban crime.

Conclusion

The STNPP model confirms that crime "flows" through the street network rather than diffusing outward in a flat plane. By leveraging GATs to visualize a "Mark Network," the framework provides interpretable insights into how different types of crimes in specific urban settings influence one another. Future work may involve incorporating directed networks (one-way streets) and dynamic covariates like weather or real-time traffic to further refine predictive power.

Summary updated

Abstract

Self-exciting point processes are widely used to model the contagious effects of crime events living within continuous geographic space, using their occurrence time and locations. However, in urban environments, most events are naturally constrained within the city's street network structure, and the contagious effects of crime are governed by such a network geography. Meanwhile, the complex distribution of urban infrastructures also plays an important role in shaping crime patterns across space. We introduce a novel spatio-temporal-network point process framework for crime modeling that integrates these urban environmental characteristics by incorporating self-attention graph neural networks. Our framework incorporates the street network structure as the underlying event space, where crime events can occur at random locations on the network edges. To realistically capture criminal movement patterns, distances between events are measured using street network distances. We then propose a new mark for a crime event by concatenating the event's crime category with the type of its nearby landmark, aiming to capture how the urban design influences the mixing structures of various crime types. A graph attention network architecture is adopted to learn the existence of mark-to-mark interactions. Extensive experiments on crime data from Valencia, Spain, demonstrate the effectiveness of our framework in understanding the crime landscape and forecasting crime risks across regions.
BibTeX
@misc{dong2024spatio,
  author = {Zheng Dong and Jorge Mateu and Yao Xie},
  title = {{Spatio-Temporal-Network Point Processes for Modeling Crime Events with Landmarks}},
  year = {2024},
  eprint = {2409.10882},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.10882},
}
Figure