One of my comp sci teachers at Cornell actually solved this problem pretty generally and implemented their research for LimeWire/Gnutella. It was called Credence, and it does just what you're saying, but crowdsourced.
Okay, so here's the basic idea: you start to vote, "I liked this, I didn't like that, I liked this." This program Credence would say, "Oh, your preferences are a lot like Alice (someone you don't know), whose preferences are also like Bob (someone who hasn't even rated anything similar to you), and you three are pretty much the exact opposite of Carol (someone else you don't know). Carol thumbs-downed X while Alice and Bob thumbs-upped it, so I'm going to suggest that it's a thumbs-up for you as well."
Notice the key elements of transitivity and negative correlation. Someone who disagrees with you all the time saying "this is bad" could actually make something good.
The idea was to kill linkspam by saying, "if spammers want to thumbs-up their own content, feel free -- but they'd better thumbs-up good content and thumbs-down other spam, at least moreso than they thumbs-up their own content, otherwise nobody will listen to them." In that sense the only way for a spammer to spam is if they contribute to killing spam in general, and it becomes a self-defeating business model.
Yeah, cluster analysis is a pretty great approach as long as you have enough users to cluster around, so often you have to have a non-social component to start things off for low amounts of users (rec more $$, Chinese, and/or San Mateo options because you liked Joy Luck Place; get more data; then rely more on social once you have data points.)
For things that mine Yelp, it's especially convenient because you can mooch data off of Yelp's social graphing. (Of course, scraping that would totally go against their ToS.)
Love the concept. Let the spammers tear each other up in a race to the bottom, while legitimate users get real value from using the system. Looking forward to reading this paper.
Okay, so here's the basic idea: you start to vote, "I liked this, I didn't like that, I liked this." This program Credence would say, "Oh, your preferences are a lot like Alice (someone you don't know), whose preferences are also like Bob (someone who hasn't even rated anything similar to you), and you three are pretty much the exact opposite of Carol (someone else you don't know). Carol thumbs-downed X while Alice and Bob thumbs-upped it, so I'm going to suggest that it's a thumbs-up for you as well."
Notice the key elements of transitivity and negative correlation. Someone who disagrees with you all the time saying "this is bad" could actually make something good.
The idea was to kill linkspam by saying, "if spammers want to thumbs-up their own content, feel free -- but they'd better thumbs-up good content and thumbs-down other spam, at least moreso than they thumbs-up their own content, otherwise nobody will listen to them." In that sense the only way for a spammer to spam is if they contribute to killing spam in general, and it becomes a self-defeating business model.
It's a pretty nifty idea. PDF here: