Pigeonholing Algorithms & Self-fulfilling Prophecies

Over the past decade, I've seen a number of online entities make attempts at categorizing my tastes. So far, Amazon's system seems to be the best, but overall, these algorithms seem to do a lackluster job and a lot of pigeonholing. They seem to make suggestions based on the results of previous clusterings.

Acknowledging that you like something from within a particular cluster seems to further reinforce the position of your tastes within the cluster and (and given how most clustering algorithms work, it probably tends to galvanize the identity of the cluster as well). A selection within some particular cluster seems to reduce the probability of system recommendations from some other cluster.

If you admit you like both R.E.M. and Idlewild, you can count on a lot of R.E.M.ish stuff, but it seems you may wind up waiting for hell to freeze over before a Cardigans recommendation comes along, because you've been marked as an fan of R.E.M.ish groups... or something like that... you get the idea. There seems to be a lot of room for improvement in this area. Too often these recommendation systems seem to translate into vicious circles of selection bias and self-fulfilling prophecies.

Update: Yahoo's Y! Unlimited music service offers a "Play Popular" songs option. If playing songs this way contributes to their popularity score, the initial popularity rankings (however they're initially determined) will simply get more and more reinforced the more the "Play Popular" choice is selected. This is another example of self-fulfilling prophecy, although in this case it is manifestly obvious. It leaves me wondering how they do handle the scoring for their popularity rankings.