Classification for predicting offender affiliation with murder victims

The National Incident-Based Reporting System (NIBRS) is used by law enforcement to record a detailed picture of crime incidents, including data on offenses, victims and suspected arrestees. Such incident data lends itself to the use of data mining to uncover hidden patterns that can provide meaningful insights to law enforcement and policy makers. In this paper we analyze all homicide data recorded over one year in the NIBRS database, and use classification to predict the relationships between murder victims and the offenders. We evaluate different ways for formulating classification problems for this prediction and compare four classification methods: decision tree, random forest, support vector machine and neural network. Our results show that by setting up binary classification problems to discriminate each type of victim–offender relationship versus all others good classification accuracy can be obtained, especially by the support vector machine method and the random forest approach. Furthermore, our results show that interesting structural insight can be obtain by performing attribute selection and by using transparent decision tree models.