this analysis on music would only tell you about what you listen to vs. what you have. in a global sense you would need a global data set. so this would conclusively answer the question: is my music library representative of what i actually like?

for instance, i think right now i’d be doing a lot better than say, back in the day when i stored and organized everything under the sun.

I definitely believe that I have low variance in my music library on a kind of global scale. My classical music is limited, I have very little jazz, and I don’t have much in the way of modern top of the pops. However within certain genres I have a great degree of variance.

It is not really an argument along the same lines. It is more of a “depth” vs. “breadth” distinction. Then of course there is a whole other system of value there.

the MUCH more complex problem is figuring out how to intelligently build those input vectors for each message. that is without question a billion times more difficult than the stuff i wrote about here. and i dont have any idea how to do it.

secondly: tags are useless when you can create arbitrary sets that represent the MAXIMUM amount of variance between groups.

I don’t understand most of the math behind it but your argument seems to be primarily that most tags are not descriptively significant in terms of differentiating between what to recommend or not (whether that be spam to be deleted or music to be listened to).

Last.fm is a social method basing recommendations on similarity actual playlists listened to. If I listen to Band A often and the users with my most similar “tastes” also listen to Band A and Band B but I don’t listen to Band B then B is recommended.

Is there an actual functional difference between the two methods or do they both break down to and arbitrary tagging system?