Testing spatially conscious machine learning models to forecast crime. A comparison case study for the prediction of property hotspots in Vienna and in Budapest.

• In Event: Advances in Quantitative Methods Panel 2: Spatial and Geographic Modeling in Criminology

Abstract

This study introduces spatially conscious machine learning (SC_ML) models for forecasting crime. SC_ML models incorporate spatial features into a machine learning modelling workflow. Machine learning (ML) algorithms are among the most used approaches to perform geographical modelling and are being used in various application domains, including criminology. In fact, over the last couple of years, numerous studies within the domain of big data policing have focused on applying ML methods to predict crime in place and time. However, such models lack spatial awareness as compared to spatial statistical models. In our study, we introduce spatial awareness by extracting spatial properties of the data and incorporating them into an ML modelling workflow with a feature engineering process. More specifically, we engineer the spatial lag that captures the spatial autocorrelation in the data and the spatial regions that capture the spatial heterogeneity in the data. Various combinations of SC_ML models are compared against traditional a-spatial ML models. Our case study involves the prediction of property crime hotspots in Vienna (Austria) and Budapest (Hungary) and offers a statistical comparison between the two cities. Hotspots of property crimes are predicted at a spatial unit of 300 by 300 meters and a monthly temporal unit for the period between 2020 and 2023. The results show the performance advantages of SC_ML models and the modelling challenges compared to the traditional ML models.

Authors

• Ourania Kounadi, Assistant Professor Department for Geography and Regional Research, University of Vienna

• Andrea Pődör, Institute of Geoinformatics, Alba Regia Technical Faculty, Óbuda University