ST3DNetCrime: Improved ST-3DNet Model for Crime Prediction at Fine Spatial Temporal Scales

Qifen Dong, Yu Li, Ziwan Zheng, Xun Wang, Guojun LiISPRS Int. J. Geo Inf 2022 · DOI

TL;DR

This paper introduces ST3DNetCrime, an enhanced deep learning model for predicting crime at fine spatial and temporal resolutions. By incorporating diurnal periodic integral mapping to handle sparse crime data and a refined spatio-temporal feature extraction mechanism, ST3DNetCrime achieves improved prediction accuracy and robustness compared to existing methods.

What Problem Does This Paper Address?

Accurate crime prediction is vital for urban development and public safety, enabling better resource allocation and crime prevention strategies. However, predicting crime at fine spatial and temporal scales (e.g., specific neighborhoods by the hour) presents significant challenges. Crime data is often sparse and irregularly distributed across both space and time, making it difficult to model the complex spatio-temporal correlations that influence criminal activity. Existing models often struggle to capture these intricate relationships, particularly the varying contributions of recent versus historical crime patterns. This paper aims to address these limitations by developing a model that can effectively learn from sparse, irregular data and capture nuanced spatio-temporal dependencies.

Methods

The core of ST3DNetCrime is an adaptation of the ST-3DNet architecture, originally designed for traffic forecasting. The authors introduce several key improvements:

  • Diurnal Periodic Integral Mapping: To combat data sparsity and irregularity, they employ a technique that integrates crime data over 24-hour periods. This process smooths out the data, making it more amenable to deep learning models and revealing daily periodicity.
  • Enhanced Temporal Feature Extraction: Instead of a single "period" component, ST3DNetCrime explicitly models three temporal properties of crime:
  • Closeness: Captures the influence of very recent crime data.
  • Period: Accounts for recurring patterns (e.g., daily or weekly cycles).
  • Trend: Considers the impact of distant historical crime data. Each of these temporal properties is processed by a dedicated component designed to extract spatio-temporal features.

  • Modified Spatio-Temporal Modeling: The model utilizes 3D convolutional layers and residual units to extract spatio-temporal features from each of the temporal components. A "Recalibration" (Rc) module is used to quantify and weigh the contribution of features from different regions, addressing the heterogeneity of spatio-temporal correlations. The outputs from the closeness, period, and trend components are then fused, with learnable parameters allowing the model to dynamically adjust the influence of each temporal property.

Datasets

The study utilizes a real-world crime dataset from Los Angeles, California. The data spans from July 1, 2015, to January 1, 2016, containing approximately 110,662 crime records with detailed time and location information. For the experiments, the researchers focused on a specific area within Los Angeles (defined by latitude and longitude ranges) where the majority of crimes occurred, resulting in 98,731 crime records. This dataset is used for both training and testing the ST3DNetCrime model and its variants.

Key Insights and Ideas

  • The Importance of Fine-Grained Temporal Patterns: The explicit modeling of closeness, period, and trend components, rather than lumping them together, significantly improves crime prediction. This suggests that recent events, recurring cycles, and long-term historical patterns all play distinct and important roles in predicting future crime.
  • Diurnal Mapping for Sparsity: The diurnal periodic integral mapping is crucial for making sparse and irregular crime data usable by deep learning models. It effectively enhances data density and reveals underlying temporal structures, leading to better model performance and robustness.
  • Robustness Through Simplicity: The study found that ST3DNetCrime exhibits good robustness and performance when using a smaller number of filters and residual units. This is a valuable insight for practical deployment, as it suggests that simpler model configurations can achieve competitive results while reducing computational costs and training time.
  • Holistic Spatio-Temporal Capture is Key: The best performance is achieved when all three temporal components (closeness, period, trend) are simultaneously utilized. This highlights that a comprehensive understanding of spatio-temporal crime dynamics requires considering a wide range of historical influences.

Key Results

  • Superior Performance: ST3DNetCrime outperformed several baseline methods, including SVR, LSTM, ConvLSTM, ST-ResNet, and the original ST-3DNet, achieving the lowest Root Mean Square Error (RMSE) of 0.307223.
  • Significant Reductions in RMSE: Compared to LSTM and ConvLSTM, ST3DNetCrime reduced RMSE by 24.5% and 22.6%, respectively. It also showed a 1.63% reduction compared to the original ST-3DNet.
  • Improved Robustness: The proposed model demonstrated better robustness, with less fluctuation in RMSE across different parameter configurations compared to variants like ST-3DNet-s.
  • Component Contributions: The closeness component was found to have the greatest impact on prediction accuracy, followed by the period component, and then the trend component, indicating that recent crime data is the most influential factor.

Limitations and Future Work

The authors acknowledge several areas for future research:

  • Incorporating Auxiliary Data: The model could be further improved by integrating additional crime-related data, such as weather patterns, public holidays, Points of Interest (POI), and public service complaints.
  • Crime Type Specificity: The current model treats all crime types uniformly. Future work could explore models that distinguish between different crime types, as their spatio-temporal patterns can vary significantly, and crime data is often sparser for specific types.
  • Dynamic Model Rebuilding: The authors suggest investigating methods to trigger model retraining or rebuilding at appropriate intervals to maintain prediction accuracy, especially in dynamic environments.
  • Fine Spatial Temporal Scales in Specific Contexts: The paper mentions the challenge of studying crime prediction in the context of ongoing public health management (e.g., anti-COVID-19 measures), suggesting this as a potential future research direction.

Summary updated

BibTeX
@article{dong2022st,
  author = {Qifen Dong and Yu Li and Ziwan Zheng and Xun Wang and Guojun Li},
  title = {{ST3DNetCrime: Improved ST-3DNet Model for Crime Prediction at Fine Spatial Temporal Scales}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2022},
  doi = {10.3390/ijgi11100529},
}
Figure