Data Base Investigation as a Ranking Problem

When data mining for forensic investigations, we are typically confronted with strongly imbalanced classes. Moreover, the labels of the non-target or negative class are usually not confirmed. In other words, the non-positive objects are unlabeled. For these situations classification methods are not well suited. We propose to approach these problems as ranking problems. We apply several supervised learning methods, including recently developed methods that are specifically aimed at optimizing ranking performance. With a true investigation dataset, we show the improvement over the prior probabilities using the ranking approach. It turns out that some two-class classification methods perform competitively on ranking performance, while the true ranking methods do not stand out.