One way around the mafia problem is a kind of normalisation, though it has its own flaws. A negative vote from someone who always gives negative votes counts for less than a negative vote from someone who generally gives positive votes. You can add in thresholding and voting ring detection for some fun, resulting in normalisation of their opinion to some degree.
There's also normalisation on the other end, how many of the total votes are negative or positive - but this doesn't really work unless there's a transaction cost, as fake accounts can shill the voting.
One of my ideas about moderation and rating systems is that the act of moderation tells us more about the person doing the moderating than it does about the story/post/person being moderated.
If we think about moderation as relational rather than objective or absolute or aggregating, we can treat the user's moderation activity as their particular web of relations to various other objects (stories/posts/users).
We could then drop all these relationships into an algorithm built to examine large matrices of relationships, say... PageRank. I wrote a tl;dr piece on moderation and online communities here: http://www.kuro5hin.org/story/2009/3/12/33338/3000
There's also normalisation on the other end, how many of the total votes are negative or positive - but this doesn't really work unless there's a transaction cost, as fake accounts can shill the voting.