Monday, 17 September 2007


The best thing about working in MIR research is that it’s part of the job to spend lots of time listening to music. Which makes me realize that I've been working very hard this weekend ;-)

I spent the last hours listening to music listening I found playing around with a recommender. It's one of those recommenders which takes one of my favorite tracks as input and returns a list of similar tracks. Seeing how amazing some of the recommendations are makes me wonder if I’ll ever again bother to browse lists of similar artists to find new music.


ben said...

The thing I find interesting here is the common assumption that high 'similarity' (content based audio, or through artist based metadata solutions, etc...) is mostly (if not completely) equivalent to a good recommendation (or a good member of a playlist). I don't necessarily think this is looking at the right sort of question. Clearly there is a relationship that leads to good recommendations which could involve broad similarity, though I tend to think it this will always be somewhat lacking. I have some other ideas (especially on the playlist front) though they are still in development. Though I will say it involves many weak classifiers and social networks...

Also, as this is relevant, I see playlists as a sort of informed (can it be assumed that you like listening to most, if not all of your local collection?) and bounded (orders of magnitude less media than the universe of digital music) special case of recommender. Though I think in the playlist problem temporal order matters more than in the recommender.

Elias said...

ben, I totally agree :-)

A good recommendation is generally not one that sounds exactly the same to a given seed song. I'm looking forward to see the outcomes of your work!

Regarding playlist generation being a form of music recommendation: I think it's worth while to distinguish them. Of course anything done by a computer can be considered a recommendation. E.g. my spell checker recommends I write playlist as two seperate words. However, I wouldn't call that a recommendation system.

Generating playlists given a user's personal collection, and recommending new music to this user are two different problems with very different challenges. Although both require the systems to be able to identify tracks that have something in common.