The general idea of a recommender system is that it asks for a few examples of things you like and then gives you more things it thinks you might like, based on its knowledge of other people’s preferences. One problem you can often run into when using a recommender system is a bias towards popular items, which are not really that close to what you like but have the favor of many users because of their high visibility.
Sebastian has some good thoughts over at Many2Many around the issue of how to apply recommendation engines to surface the Long Tail of content. We've been thinking about this a lot as it is applied in the realm of television and open video publishing. Clearly, popularity and collaborative filtering based approaches suffer from surfacing too much popular/head content, so approaches to content recommendations that can readily surface things relevant to an individual based on limited interactions as well as deep knowledge of content seems to be the key. Right now, out television platforms, even with systems like TiVo's recommendations, are crude at best at surfacing relevant content to users. The answer, of course, is not just in finding the right technical approach, but it's clever application of editorial, community and consumer participation.