A Recommendation Engine to Aid in Identifying Crime Patterns

Police investigators are routinely asked to search for and identify groups of related crimes, known as patterns. Investigators have historically built patterns with a process that is manual, time-consuming, memory based, and liable to inefficiency. To improve this process, we developed a set of three supervised machine-learning models, which we called Patternizr, to help identify related burglaries, robberies, and grand larcenies. Patternizr was trained on 10 years of manually identified patterns. Problematic administrative boundaries and sensitive suspect attributes were hidden from the models. In tests on historical examples from New York City, the models perfectly rebuild approximately one-third of test patterns and at least partially rebuild approximately four-fifths of these test patterns. The models have been deployed to every uniformed member of the New York City Police Department through a custom software application, allowing investigators to prioritize crimes for review when building a pattern. They are used by a team of civilian crime analysts to discover new crime patterns and aid in making arrests.