The real problem with that is that the concept of similar is pretty hard to define and is always implemented poorly. Just because I like Pearl Jam doesn't mean that I like Creed in any way. Quality (which is subjective anyway) seems never to be taken into account in such algorithms.

Well I don't think computers can figure it out to that level. But if you have more characteristics about each song, (like the AllMusic database, for instance) and can find patterns in those, you'll at least have better results than simply throwing random songs at the user. Tying in information on what other people do (like Amazon's "customers who bought Pearl Jam also bought Creed" feature) helps a little bit too. It's not going to be perfect, but all things being equal, an imperfect system is better than the absence of any system at all.

Or that's what we're banking on in our paper, anyway.