I think @jayrhacker is suggesting using such techniques to predict “troll” or “not troll” given the posting history/removed comments/etc. My personal thought is that whatever system replaces karma, it should be understandable to the typical user. I think its possible Bayesian inference could be used in developing the system, but the end system should be explainable without it.