One thing that may work well here is to train two models and ask which fits the data better (likelihood). Or, to leverage the full weight of the data in traini

One thing that may work well here is to train two models and ask which fits the data better (likelihood). Or, to leverage the full weight of the data in training submodels and then leverage that in the classifiers. Fraud is particularly difficult though, because the entities are actively trying to thwart your attempts to detect them. Outlier detection is a must certainly. Collective entity resolution helps a lot too. It might even be worth seeing if you can use LDA to cluster the fraudsters together. It's just not a domain suited to simple models or a singular approach.