Khalfa Robin, Theinert Naomi, Hardyns Wim
Department of Criminology, Criminal Law and Social Law, Ghent University, Universiteitstraat 4, Ghent, 9000 Belgium.
Faculty of Social Sciences, University of Antwerp, Sint-Jacobstraat 2, Antwerp, 2000 Belgium.
Comput Urban Sci. 2025;5(1):27. doi: 10.1007/s43762-025-00185-x. Epub 2025 May 9.
This study empirically compares multiple eXplainable Artificial Intelligence (XAI) techniques to interpret short-term (weekly) machine learning-based burglary predictions at the micro-place level in Ghent, Belgium. While previous research predominantly relies on SHAP to interpret spatiotemporal crime predictions, this is the first study to systematically evaluate SHAP alongside other XAI techniques, offering both global and local model interpretability within the context of crime prediction. Using data from 2014 to 2018 on residential burglary, repeat and near-repeat victimization, environmental features, socio-demographic indicators, and seasonal effects, we trained an XGBoost model with 76 features to predict weekly burglary hot spots. This model serves as a basis for comparing the interpretative power of different XAI techniques. Our results show that built environment and land use characteristics are the most consistent global predictors of burglary risk. However, their influence varies substantially at the local level, revealing the importance of spatial context. While global feature importance rankings are broadly aligned across XAI techniques, local explanations, especially between SHAP and LIME, often diverge. These discrepancies highlight the need for careful method selection when translating predictions into crime prevention strategies. In addition, this study demonstrates that short-term burglary risks are influenced by complex interactions and threshold effects between environmental and social disorganization features. We interpret these findings through the lens of criminological theory, and argue for more integrated approaches that go beyond examining the isolated effects of specific crime predictors. Finally, we call for greater attention to the methodological implications that arise from applying different interpretability techniques, particularly when machine learning model outputs are used to inform crime prevention and policy decisions.
The online version contains supplementary material available at 10.1007/s43762-025-00185-x.
本研究通过实证比较多种可解释人工智能（XAI）技术，以解读比利时根特微观层面基于机器学习的短期（每周）入室盗窃预测。虽然先前的研究主要依靠SHAP来解读时空犯罪预测，但这是第一项系统评估SHAP以及其他XAI技术的研究，在犯罪预测背景下提供了全局和局部模型可解释性。利用2014年至
2018年期间关于住宅入室盗窃、重复和近乎重复受害情况、环境特征、社会人口指标以及季节效应的数据，我们训练了一个具有76个特征的XGBoost模型来预测每周的入室盗窃热点地区。该模型作为比较不同XAI技术解释力的基础。我们的结果表明，建成环境和土地利用特征是入室盗窃风险最一致的全局预测因素。然而，它们在局部层面的影响差异很大，揭示了空间背景的重要性。虽然全局特征重要性排名在不同XAI技术之间大致一致，但局部解释，尤其是SHAP和LIME之间的解释，往往存在分歧。这些差异凸显了在将预测转化为犯罪预防策略时仔细选择方法的必要性。此外，本研究表明，短期入室盗窃风险受环境和社会无序特征之间复杂的相互作用和阈值效应影响。我们通过犯罪学理论的视角解读这些发现，并主张采用超越考察特定犯罪预测因素孤立效应的更综合方法。最后，我们呼吁更多关注应用不同可解释性技术所产生的方法学影响，特别是当机器学习模型输出用于为犯罪预防和政策决策提供信息时。
在线版本包含可在10.1007/s43762-025-00185-x获取的补充材料。
