My first inclination would be "bot-like" behavior, like rapid voting outside of human response times, posting identical comments to multiple threads, easy things like that.

It would get harder to edge into the cloud of data for more precise definitions, but some of the more obvious behavior may be a starting point.

I realize it isn't an easy classification problem, but there are some edges where it could be attacked from even in a limited sense.