Well, the goal of these systems isn't to be perfect... You'll never match everyone's tastes. All they're trying to do is give you recommendations and suggestions that are statistically going to make more "hits" (which equals more purchases if you decide you like it) and correlating your collection with others and using their preferences to guide your recommendations is one of the better ways to do this. There are several projects I've found which attempt to match songs by "musical similarity" (including this one) but I would argue that using ONLY this kind of metric (and excluding population-based data) would tend to limit one's exposure to new music which may be somehow relevant to songs they already own, but might not necessarily "sound" like it. While Soundgarden and Pearl Jam might not correlate highly via a musical similarity algorithm, the fact that you see them both in a lot of peoples' collections *should* be a factor in any recommendation engine... Just not the only factor.

Right now my project team and I have four basic factors that are going to go into the equation which will rank songs for new music recommendation. Population-based data like MusicMatch is using is one of those factors, and right now I think we're leaning towards it being 30% of the equation by default. But we've got seven weeks to figure out if that's good or not, and really, since we won't end up with a legitimate working system, it's all a bit hypothetical. But if I were developing a real recommendation engine, that's about as high as I'd go with the use of population-based data.