Interpretable spatial machine learning for understanding spatial heterogeneity in factors affecting street theft crime

Abstract

Machine learning techniques, such as random forest (RF), are increasingly utilized in geographical studies, including crime analysis. While RF excels in predictive accuracy and reduces overfitting, it does not account for spatial heterogeneity, where relationships between crime and its determinants vary by location. To address this limitation, this study utilizes a geographically weighted random forest (GWRF), which breaks down RF into local sub-models. A comprehensive set of variables is integrated to provide a thorough understanding of the factors influencing street theft crimes, including risk populations, streetscape environments, social disorganization, crime attractors and generators, and transportation accessibility. The results of the machine learning analysis are interpreted through an interpretability framework. Findings indicate that GWRF achieves greater predictive accuracy than traditional methods, reaching up to 80% accuracy on unseen data. Notably, the on-street population, derived from street view images, is the most significant contributor to street theft. Additionally, the results facilitate the examination of spatial non-stationarity concerning the importance of explanatory variables. GWRF effectively bridges machine learning and geographical models, enhancing our understanding of street theft occurrences. The local importance results enable the development of targeted safety interventions, rather than relying on general guidelines applicable to the entire region.