I'd like to see a recommendation program that trains on the many, many aspects of music itself that a person listens to. Rather than artists/labels/style slots/norms/years of release/charts ... it'd analyze your preferences in tempo, melody, harmony, instruments (type and number), vocal styles, novelty. (Getting started would take a while, say, over a month.)

Then to refine its model, it'd have you rate some picks (old and new based on what it knows so far), and analyze how your responses fit its model. Would it get better faster for people who like a wide variety, or those with a few specialities?