What makes a good recommendation?

I have been using the personalized on-line radio service Pandora for the past few weeks and enjoying it, but only heard about Last.fm while reading this wonderful weblog post contrasting the recommendation algorithms behind the two systems. The post as a whole is a great read, but I particularly like its discussion of how context is relevant to selecting the superior technology. I would love to see a study that actually considered whether a collaborative filter or an attribute analysis approach was more effective in the context of music. Anecdotally, I’m going to try Last.fm out for a week or two and see what seems to work better for me. Krause found that for him, Last.fm was better, but suggested that Pandora holds more promise for incorporating some of the advantages of Last.fm and continuing to improve.
Of course, Krause makes the key point early in his post: “better algorithms are nice, but better data is nicer”. For music recommendations, the system that heppens to know more about the music you happen to like will probably be the best system for you.