I once had the idea of using markov chains to help build playlists. I find the order of songs can be very important, and that way I could build a more random playlist based liking song B after song A, and just keep going.

The problem was two fold, one, I really love whole albums when done right. So that most of the music I listen to is album based. The markov chain is pointless then. Two, I would need a really big data set to make it worthwhile. Rdio has that data set (or possibly even Echo Nest).