I think the hardest problem with that approach would be avoiding a false-positive of detecting people who hold controversial opinions as trolls.
The moderation graph between users and comments would be really interesting. There's a lot of information in there to be mined.