Comparing machine learning-based crime hotspots vs police districts: what’s the best approach for crime forecasting?

Criminal activity poses a significant challenge in urban environments, impacting public safety, economic stability, and overall quality of life. As a result, the efficient allocation of public security resources based on spatio-temporal crime prediction models has become a critical concern for urban management. To this purpose, many crime forecasting approaches first split city territories into partitions based on crime rates and trends, with each partition reflecting criminal dynamics of its specific area. Then, crime forecasting models are extracted for each area, to monitor and predict how crime rates evolve over time within each partition. However, traditional spatial partitioning approaches, which divide cities into predefined police districts based on geographic and operational considerations, often fail to account for variations in crime patterns. In contrast, machine learning-based approaches could dynamically adapt to areas with differing crime frequencies and densities, making them particularly effective in cities characterized by diverse population distributions and crime activity levels. This study examines the impact of various partitioning techniques on crime forecasting performance, comparing the traditional static division of the city into police districts with machine learning approaches, specifically density clustering algorithms, for detecting crime hotspots. The experimental evaluation, conducted on two real-world case studies, i.e. Chicago and Los Angeles crime data, demonstrates the effectiveness of density-based clustering in identifying multi-density crime hotspots. Compared to traditional police district partitioning, these data-driven methods offer significant advantages in improving crime forecasting accuracy across urban environments.