[Paper Review] Deep Learning for Real-Time Crime Forecasting and its Ternarization

Bao Wang, Penghang Yin+4 more|arXiv (Cornell University)|Nov 23, 2017

Anomaly Detection Techniques and ApplicationsComputer Science28 references20 citations

TL;DR

This paper proposes a spatial-temporal residual network (ST-ResNet) for real-time crime forecasting at the hourly scale in Los Angeles, using a novel crime data representation that enhances sparsity through temporal cumulative aggregation and spatial super-resolution. The model achieves high accuracy with minimal performance loss when fully ternarized, reducing memory and energy use for deployment on resource-constrained devices.

ABSTRACT

Real-time crime forecasting is important. However, accurate prediction of when and where the next crime will happen is difficult. No known physical model provides a reasonable approximation to such a complex system. Historical crime data are sparse in both space and time and the signal of interests is weak. In this work, we first present a proper representation of crime data. We then adapt the spatial temporal residual network on the well represented data to predict the distribution of crime in Los Angeles at the scale of hours in neighborhood-sized parcels. These experiments as well as comparisons with several existing approaches to prediction demonstrate the superiority of the proposed model in terms of accuracy. Finally, we present a ternarization technique to address the resource consumption issue for its deployment in real world. This work is an extension of our short conference proceeding paper [Wang et al, Arxiv 1707.03340].

Motivation & Objective

• Address the challenge of accurate real-time crime forecasting at fine spatial-temporal scales, particularly given sparse and stochastic crime data.
• Overcome limitations of existing classification-based models by modeling crime intensity as a regression problem for more nuanced predictions.
• Improve model efficiency for real-world deployment by introducing ternarization to reduce memory and computational costs without significant accuracy loss.
• Develop a robust data representation that enhances weak signals in sparse crime data through temporal and spatial signal enhancement techniques.

Proposed method

• Represent crime data as spatial-temporal images where each pixel encodes the number of crimes in a neighborhood grid at a given hour.
• Apply diurnal cumulative crime aggregation in the temporal dimension to amplify predictable patterns and reduce noise.
• Use bilinear interpolation for spatial super-resolution to enhance spatial granularity and capture local crime dynamics.
• Adapt the ST-ResNet architecture to model complex spatiotemporal dependencies in crime patterns using residual learning.
• Implement a ternary weight quantization scheme where weights are constrained to {-1, 0, +1} with a learnable scaling factor to reduce model size and energy consumption.
• Use a two-stage training process: first train full-precision weights via ADAM, then project them to ternary values using a projection function to preserve performance.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively forecast crime intensity at the hourly scale in micro-geographic regions despite sparse and stochastic data?
• RQ2 How does the proposed data representation—combining temporal cumulative aggregation and spatial super-resolution—improve forecasting accuracy compared to raw crime data?
• RQ3 To what extent does ternarization of ST-ResNet weights reduce model size and inference cost while maintaining predictive accuracy?
• RQ4 How does the regression-based ST-ResNet model compare to classification-based approaches in capturing the true distribution of crime events?
• RQ5 Can the model be effectively deployed on mobile or edge devices with limited computational resources through quantization techniques?

Key findings

• The proposed ST-ResNet model achieves a test RMSE of 0.207 on the full-precision version, demonstrating high accuracy in forecasting crime intensity at the hourly scale.
• The fully ternarized ST-ResNet model incurs only a small performance degradation, with a test RMSE of 0.242, showing minimal accuracy loss despite 32× memory reduction.
• Ternarization reduces model size and accelerates inference by replacing 32-bit floating-point weights with 1-bit ternary values and a single scaling factor.
• The combination of temporal cumulative aggregation and spatial super-resolution significantly enhances signal strength in sparse crime data, improving model generalization.
• The regression-based approach outperforms binary classification models by capturing the full distribution of crime counts rather than just presence/absence.
• The model is scalable and suitable for real-time deployment in law enforcement, offering actionable forecasts for proactive patrol planning.