Abstract
Homicide prediction is a challenging task due to the spatio-temporal sparsity of these crime events. In this paper, we report the results of using several approaches to mitigate this sparsity condition in machine learning models specially tailored towards modeling homicide events. Since spatial resolution is a direct determinant of sparsity, we focus on the performance of these models across different resolutions of interest to police authorities. We use a simple count model as a benchmark and propose enhancements directed towards improving prediction performance. We then compare the results to more complex models motivated by manifold learning and graph signal processing methods. We found that the simple benchmark models are as good as state-of-the-art models for low resolution, but, as resolution increases, machine learning models outperform the benchmark. These results provide a rationale for the use of state-of-the-art machine learning models for homicide prediction at the high resolution required for the deployment of police resources.
Introduction: The Sparsity Challenge
In the field of predictive policing, homicides represent a significant data science challenge. Unlike high-volume crimes such as theft, homicides are statistically rare, creating a condition known as spatio-temporal sparsity. When a city is divided into a grid for analysis, the vast majority of cells contain zero events for any given time period.
As police departments require higher spatial resolution to deploy patrols effectively—moving from 1,000m blocks to precise 150m cells—this sparsity problem intensifies. Most traditional statistical models struggle when faced with an abundance of zeros. This study investigates how different modeling strategies perform as spatial resolution increases and the data becomes thinner.
Data and Spatial Resolution
The study utilized a high-quality dataset from Bogotá, Colombia, specifically the SIEDCO dataset, which includes daily homicide records from January 2018 to December 2019. The data was split into 78 weeks for training and 26 weeks for testing.
As shown in the spatial discretization analysis, moving from a 1,000m grid to a 150m grid creates a map of over 18,000 cells, most of which have no recorded history of violence. The temporal distribution of these events also highlights the difficulty of the task, as shown in the training and testing split below.
Modeling Strategies
The researchers compared three tiers of models to address this sparsity:
- Baseline Count Models: These range from "Static Count" models to versions using sliding windows and time decay factors to weight recent events more heavily.
- Augmented Models: These incorporated "street fight" data as a proxy, leveraging the fact that roughly 80% of homicides in Bogotá are linked to such altercations.
- State-of-the-Art (SOTA) ML Models:
- Kernel Warping (KW): This treats the city as a "manifold" (a curved mathematical surface) shaped by the density of historical crimes and street fights, allowing it to predict homicides in areas with no prior history.
- Graph Laplacian of Gaussian (GLoG): This treats the city as a graph where cells are nodes, using signal processing to detect abrupt changes in crime signals and identify emerging hotspots.
Key Results and Insights
The study revealed a "natural limit" to simple counting. If a homicide occurs in a cell with zero prior events, a count model will inevitably miss it. As resolution increases, more homicides fall into these "zero-history" cells, causing baseline performance to degrade.
Key findings include:
- Resolution Sensitivity: Simple models are robust at low resolutions (large areas), but their performance plummets as the grid zooms in.
- The High-Res Winner: At the finest resolution (150m), Kernel Warping showed the most stable and superior performance. Interestingly, the performance of the Kernel Warping model actually improved across several metrics as resolution became finer, whereas baseline models degraded.
- Stability through Proxies: Models incorporating street fight data remained much more stable across resolutions, proving that auxiliary data is a vital weapon against sparsity.
Conclusion
While simple historical count models are sufficient for coarse spatial resolutions, sophisticated machine learning approaches like Kernel Warping and Graph Signal Processing are essential for the high-resolution predictions required for effective police patrolling. Future work may explore daily or shift-based predictions, though this will require new architectures to handle the even more extreme "ultra-sparse" environments that result from finer temporal scales.
Summary updated