Using machine learning to conduct crime linking of residential burglary

Abstract

Traditional crime linkage methods face challenges with complex datasets, arguably necessitating more sophisticated analytical tools. This research investigates this issue by exploring the application of machine learning, specifically the Random Forest algorithm, as a method to enhance crime linkage analysis of residential burglary cases.

Using a dataset of 200 pairs of linked residential burglaries from the United Kingdom, this study employs the Random Forest technique to examine 67 identified crime features, including those within categories related to inter-crime distance, temporal patterns, such as time and day of the week, target selection, entry behaviour, crime scene conduct, and property stolen.

The key objective is to identify and reduce predictive characteristics that reliably link burglaries, whilst potentially overcoming the limitations of conventional approaches. Findings generally support existing literature but provide increased nuance by indicating that certain factors specifically related to shorter inter-crime distances, the time and date of the offences, and the target's dwelling type, significantly contribute to accurately linking crimes. We discuss these findings in the context of existing research on the subject.

Finally, we consider the benefits of using this novel methodology as a tool for crime linking. We argue that the improved accuracy, interpretability, and provision of multiple decision trees offers significant advantages for refining crime linkage practices, both operationally and in criminological research.