The subject of this post is a required question for one of my current subjects at university.
We are to explain why we’re a member of a particular social media site, and what we get out of it.
I was recommended rdio by b, who was finding it easier to search for individual 80s tracks than in iTunes. Being stubborn and not easily swayed into a new environment it took me some time to succumb — however now that I have I find it invaluable.
What rdio does is attempt to understand your tastes, and provide you wish new musical experiences you may not have thought of. With me, the algorithm works; with b, not so much. For me, the more I use it, the better the algorithm can predict something I would like to listen to. For b, the longer he uses it, the more confused the poor program becomes.
I think the reason that it works better for me than for b is because of how I believe the algorithm works. It appears to be a statistical based model; if user a likes song x, he is likely to enjoy the same songs as other people who like song x. This is further refined by the rest of your music collection; the more correlations found between your tastes and the tastes of the other people who like song x the closer the recommendations will be to your actual tastes.
There is some tuning available to this idea, there are currently 5 stages of similarity to your own collection, or artist (previously there were 7):
With someone like me who likes music based on the sound of a group, and tends to like most of what they produce; an algorithm can predict what I would like relatively easily using a statistical model. The more similarity there are between the bands in various users collections, the more they are likely to enjoy a band that does not appear in their own collection.
For someone like b who likes individual tracks, the algorithm does not work quite so well; it is hard to build up a statistical model for individual tracks as they can differ greatly even on a single album, and when ‘Various Artist’ albums are brought into the mix it makes it even harder for a statistical model to build up. This can easily be seen with the autoplay feature.
Take the song ‘Eleanor Rigby’ for example; If one went by the name of the track alone it could be reasonably safe to assume that I would like the rest of the Beatles catalogue, however the version I actually like is by Godhead and sounds very little like the original. Likewise, listening to 2Cellos play Thunderstruck has as much chance of giving the user more cello music as it does ACDC tracks.
I use that track as an example as I would not have even heard of it without rdio, it has expanded my musical universe considerably.