Interpreting short-term burglary predictions: A comparative evaluation of XAI techniques

• In Event: Predicting crime in space and time: Examining and understanding input, throughput and output

Abstract

This study evaluates and compares different eXplainable Artificial Intelligence (XAI) techniques for interpreting short-term (weekly) burglary predictions at the micro-place level (200 by 200 meter grid cells) in Ghent, Belgium. While previous research has mainly focused on SHAP for explaining spatiotemporal crime predictions, this is the first study to systematically compare SHAP with other XAI methods in crime prevention, assessing both global and local model interpretability. Given SHAP’s computational demands, the study guides researchers in selecting suitable techniques and examines their impact on model transparency and theoretical relevance.

Using data on residential burglary and 76 predictive features—including crime opportunity, socio-demographic, economic, and seasonal factors from 2014-2018—XGBoost models are trained to predict weekly burglary hot spots. This serves as a baseline to compare the different XAI techniques across different levels of model interpretation, including global feature contributions, feature effects and interactions and local explanations.

Findings show that built environment and land use are the strongest global predictors, but their influence varies locally, underscoring the importance of spatial context and processes. While global feature importance scores align across XAI techniques, local-level explanations diverge substantially, particularly between SHAP and LIME. These discrepancies highlight the need for careful method selection in crime prediction. More broadly, results indicate that short-term burglary risks are influenced by complex interactions between environmental and social factors rather than isolated feature effects.

Future work should explore how discrepancies between XAI techniques affect decision-making in crime prevention and refine methodological approaches to better capture urban crime dynamics. This research is part of the Big Data Policing (BIGDATPOL) program (ERC, BIGDATPOL, 101088156), which advances data-driven crime analysis.

Authors

• Robin Khalfa, Ghent University

• Naomi Theinert, Ghent University

• Wim Hardyns, Ghent University, Department of Criminology, Criminal Law and Social Law