FWIW, I don't think anyone knows, including the people in charge.
This type of system is a typical AI problem, where you put in a bunch of data with known "bad actors" and try to predict the next bad actor.
Silly example: It could be that the system decides that people who's family name starts F and are born in August and are Canadian or Madagascarian of origin. are more likely to be bad actors, but fundamentally the system is as good as the person writing it and I doubt anyone has any way or incentive to make it better.