Uncertainty-Aware Crime Prediction With Spatial Temporal Multivariate Graph Neural Networks

Zepu Wang, Xiaobo Ma, Huajie Yang, Weimin Lyu, Yang Liu, Peng Sun, Sharath Chandra GuntukuICASSP 2025 · DOI

TL;DR

This paper introduces STMGNN-ZINB, a novel graph neural network framework that tackles the challenge of predicting urban crime by explicitly modeling data sparsity and uncertainty. By integrating a Zero-Inflated Negative Binomial distribution with spatial-temporal graph convolutions, the model achieves more accurate and reliable crime forecasts, offering valuable insights for urban planning and resource allocation.

What Problem Does This Paper Address?

Predicting crime is crucial for urban safety and resource management. However, crime data is notoriously sparse, especially at a granular level (small regions, specific time periods), and often exhibits a "zero-inflated" nature (many instances of zero crime). Traditional deep learning models, particularly those assuming Gaussian distributions for predictions, struggle with this sparsity and overdispersion. This can lead to inaccurate forecasts and unreliable confidence intervals, hindering effective decision-making for law enforcement and urban planners. The paper aims to bridge this gap by developing a model that can accurately capture and quantify the uncertainty inherent in sparse, discrete crime data.

Methods

The core of the proposed method, STMGNN-ZINB, lies in its novel combination of existing techniques and a new probabilistic modeling approach:

  • Spatial-Temporal Graph Convolutional Networks (STGCNs): The paper leverages DGCNs (Diffusion Graph Convolutional Networks) to capture spatial dependencies between different urban regions. It also incorporates Multivariate-Temporal Convolutional Networks (MTCNs) to model temporal correlations across different crime types within a given region.
  • Zero-Inflated Negative Binomial (ZINB) Distribution: This is the key innovation for handling data sparsity. Unlike standard models that assume continuous or Gaussian distributions, ZINB explicitly accounts for an excess of zero counts and models the count data using a negative binomial distribution. This allows the model to better represent the discrete and sparse nature of crime incidents.
  • Fusion Mechanism: The spatial embeddings from DGCNs and the multivariate-temporal embeddings from MTCNs are fused using a Hadamard product. This combined representation is then used to parameterize the ZINB distribution, enabling joint prediction and uncertainty quantification.
  • Negative Log-Likelihood Loss: To circumvent issues with KL-divergence calculations when dealing with zero values, the authors opt for a negative log-likelihood loss function, which is more robust for sparse, zero-inflated data.

Datasets

The paper evaluates STMGNN-ZINB on two real-world urban crime datasets:

  • NYC-Crimes: Covers crime incidents in New York City from January 2014 to December 2015.
  • CHI-Crimes: Covers crime incidents in Chicago from January 2016 to December 2017.

For both datasets, the cities were divided into a spatial grid (3km x 3km), resulting in 256 regions for NYC and 168 regions for CHI. Predictions are made at a daily resolution. The datasets were split into training (7:1 ratio) and testing sets, with the last 30 days of training used for validation.

Key Insights and Ideas

  • The Importance of Probabilistic Modeling for Sparse Data: The paper strongly emphasizes that assuming Gaussian distributions for crime data is fundamentally flawed due to its inherent sparsity and zero-inflation. The ZINB distribution provides a much more appropriate and interpretable way to model such data.
  • Interpretable Uncertainty: The sparsity parameter ($\pi$) of the ZINB distribution has a direct physical interpretation: the likelihood of a region being devoid of crime. This offers valuable insights for understanding crime patterns and can aid in strategic resource allocation.
  • Holistic Evaluation is Crucial: The authors highlight that relying on a single metric can be misleading. For instance, STMGNN-G performed well on KL Divergence but poorly on F1-score for the NYC dataset, demonstrating the need to assess models across various metrics (point estimation, uncertainty quantification, and discrete metrics).
  • Spatial Locality in Crime: The heatmap visualization of the sparsity parameter ($\pi$) confirms the expected spatial locality of crime, with crime activities varying gradually across regions and being more concentrated in downtown areas compared to rural outskirts.

Key Results

  • Superior Prediction Accuracy: STMGNN-ZINB consistently achieved the lowest Mean Absolute Error (MAE) across both datasets, indicating more accurate point predictions compared to baseline methods.
  • Improved Uncertainty Quantification: The model demonstrated lower Mean Prediction Interval Width (MPIW), suggesting more precise uncertainty estimates. The Prediction Interval Coverage Probability (PICP) was also competitive, aiming for the desired 90% coverage.
  • Effective Handling of Sparsity: STMGNN-ZINB outperformed other models in metrics like True Zero Rate and F1 Score, showcasing its ability to accurately capture and predict the sparse nature of crime data.
  • Best KL-Divergence: When excluding a less suitable baseline (STMGNN-G), STMGNN-ZINB showed the best performance in approximating the actual data distribution, as indicated by KL-Divergence.

Limitations and Future Work

  • Limitations: The paper acknowledges that while STMGNN-G performed well on KL Divergence and True Zero Rate for the NYC dataset, its low F1-score indicates struggles with predicting actual crime occurrences, especially in highly sparse scenarios. This underscores the challenge of balancing different aspects of prediction and uncertainty.
  • Future Work: The authors suggest that the interpretable sparsity parameter ($\pi$) can be a valuable tool for social security decision-makers to efficiently allocate limited resources. Further exploration of how this parameter can directly inform policy and intervention strategies would be a logical next step.

Summary updated

BibTeX
@inproceedings{wang2025uncertainty,
  author = {Zepu Wang and Xiaobo Ma and Huajie Yang and Weimin Lyu and Yang Liu and Peng Sun and Sharath Chandra Guntuku},
  title = {{Uncertainty-Aware Crime Prediction With Spatial Temporal Multivariate Graph Neural Networks}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10889685},
}
Figure