[Paper Review] Deep Learning for Real Time Crime Forecasting

Bao Wang, Duo Zhang+3 more|arXiv (Cornell University)|Jul 9, 2017

Anomaly Detection Techniques and Applications19 citations

TL;DR

This paper proposes a two-stage deep learning framework, ST-ResNet adapted for crime forecasting, to predict spatio-temporal crime distributions in Los Angeles with high accuracy. By applying spatial and temporal regularization and using hierarchical residual convolutional units, the model achieves strong predictive performance over a six-month period, demonstrating significant potential for real-time public safety applications.

ABSTRACT

Author(s): Wang, Bao; Zhang, Duo; Zhang, Duanhao; Brantingham, P; Bertozzi, Andrea L | Abstract: Accurate real time crime prediction is a fundamental issue for public safety, but remains a challenging problem for the scientific community. Crime occurrences depend on many complex factors. Compared to many predictable events, crime is sparse. At different spatio-temporal scales, crime distributions display dramatically different patterns. These distributions are of very low regularity in both space and time. In this work, we adapt the state-of-the-art deep learning spatio-temporal predictor, ST-ResNet [Zhang et al, AAAI, 2017], to collectively predict crime distribution over the Los Angeles area. Our models are two staged. First, we preprocess the raw crime data. This includes regularization in both space and time to enhance predictable signals. Second, we adapt hierarchical structures of residual convolutional units to train multi-factor crime prediction models. Experiments over a half year period in Los Angeles reveal highly accurate predictive power of our models.

Motivation & Objective

• To address the challenge of real-time crime forecasting due to the sparsity and irregularity of crime patterns in space and time.
• To improve predictive accuracy by enhancing weak signals in raw crime data through spatial and temporal regularization.
• To develop a scalable deep learning model capable of capturing complex, low-regularity crime distributions across multiple spatio-temporal scales.
• To evaluate the performance of the proposed model on real-world data from Los Angeles over a six-month period.

Proposed method

• Applying spatial and temporal regularization to raw crime data to amplify predictable signals and reduce noise.
• Adapting the ST-ResNet architecture, a state-of-the-art spatio-temporal deep learning model, for crime prediction tasks.
• Employing hierarchical residual convolutional units to model complex spatial and temporal dependencies in crime patterns.
• Training the model on a multi-factor representation of crime data to capture diverse influencing factors.
• Using a two-stage pipeline: first preprocessing the data, then training the deep learning model on the enhanced data.
• Leveraging residual learning to stabilize training and improve generalization in low-regularity crime data.

Experimental results

Research questions

• RQ1 Can deep learning models effectively forecast crime distributions in real time despite their sparsity and irregular patterns?
• RQ2 How does spatial and temporal regularization improve the predictability of crime data in a deep learning framework?
• RQ3 To what extent can hierarchical residual convolutional networks model complex spatio-temporal crime dynamics in urban environments?
• RQ4 How does the proposed model perform on real-world crime data from Los Angeles over a six-month period?

Key findings

• The proposed model demonstrates highly accurate predictive power for crime distribution across the Los Angeles area.
• Spatial and temporal regularization significantly enhances the signal-to-noise ratio in raw crime data, improving model performance.
• The use of hierarchical residual convolutional units enables effective learning of complex, non-linear spatio-temporal crime patterns.
• The model achieves strong predictive accuracy over a six-month evaluation period, indicating robustness and scalability.
• The framework successfully captures low-regularity crime distributions across varying spatio-temporal scales.
• The two-stage approach—data preprocessing followed by deep learning modeling—proves effective for real-time crime forecasting.