Comparing Machine Learning-Based Crime Predictions Across Micro-Geographic Units: Street Segments, Rectangular Grids, and Hexagonal Grids

This study examines the potential of alternative micro-geographic units of analysis compared to the widely used rectangular grid for predicting (monthly) micro-geographic crime risks using a machine learning approach. Specifically, this study compares the prediction performance of machine learning models (XGBoost) in deriving monthly micro-geographic risk predictions for three crime types across rectangular and hexagonal grids of varying resolutions, as well as street segments, using the hit rate, precision, and F1-score as key performance measures. Police-registered data on residential burglary, aggressive theft, and battery (2013–2018), along with environmental and seasonal data on crime predictors were used to train the models and evaluate their performance across different units of analysis and performance measures. Results show that street segments generally achieve higher hit rates compared to grid-based units, but only marginally when compared to high-resolution grids (0.0025 km²). This study thus finds no clear advantage of street segments over small grids in terms of model hit rate. In addition, using street segments and small grids comes at the cost of lower model precision, resulting in more false positive predictions. Grids with resolutions from 0.04 km² to 0.25 km² offer a more balanced performance. Further, no substantial differences were found between rectangular and hexagonal grids, indicating grid shape does not affect prediction performance. Future work should explore how model performance should be defined and operationalised within the context predicting crime risks at specific micro-geographic levels and what the implications are of employing specific micro-geographic units of analysis within the context of crime prevention.