So for something like "fraudsters don't use capital letters," does your system discover a fact like this automatically, or do I have to think up these indicators myself and hope they are relevant?
This was discovered automatically. That's one benefit of a machine learning-based system: you feed it a lot of data and tell it when fraud actually occurred and it adapts its rules predict fraud accordingly.
There is no relevancy here to be had, these are simply trivially interesting trends.