Comparing Soft Classification Methods for the Rare Type Match Problem

In this work, we compare five different methods proposed in forensic statistics to cope with the rare type match problem. This problem arises when the DNA profile of a suspect coincides with the profile from a crime sample, but it is not present in the available database collected from the population of reference. We compare the methods designed to evaluate the likelihood ratio in this framework by using a set of supervised cases and by considering each method as a classifier that provides the posterior probabilities of two alternative hypotheses, those of the prosecution and the defense, starting from a grid of prior probabilities. We compare them using the value of the posterior cross entropy and decompose it into two terms quantifying their calibration and refinement loss.