Predicting the Offender: Frequency versus Bayes

Abstract

In this paper two Bayesian approaches and a frequency approach are compared on predicting offender output variables based on the input of crime scene and victim variables. The K2 algorithm, Naïve Bayes and frequency approach were trained to make the correct prediction using a database of 233 solved Dutch single offender/single victim homicide cases and validated using a database of 35 solved Dutch single offender/single victim homicide cases. The comparison between the approaches was made using the measures of overall prediction accuracy and confidence level analysis. Besides the comparison of the three approaches, the correct predicted nodes per output variable and the correct predicted nodes per validation case were analyzed to investigate whether the approaches could be used as a decision tool in practice to limit the incorporation of persons of interest into homicide investigations. The results of this study can be summarized as: the non-intelligent frequency approach shows similar or better results than the intelligent Bayesian approaches and the usability of the approaches as a decision tool to limit the incorporation of persons of interest into homicide investigations should be questioned.