last.fm is an interesting service, it monitors what music you listen to and claims it can recommend you new music you'll like, based on your taste. I've been using it for a while now. I think it's got an OK amount of data about what I listen to, since it's got a massive list of tracks from me. The recommendations have been a bit hit-and-miss.
First the negative. Here are some of the disappointing connections:
- While I was listening to Blur, it listed similar artists as Oasis and a couple of other well-known britpop bands (The Verve, Pulp, Radiohead). Those aren't "similar", they're just part of the same milieu. Oasis and Blur might both be guitary indie-rock from the late 90s but beyond that they're musically so different, in the chords, the melodies, the lyrics, and the overall sound. A decent "similar artists" listing would find some bands that were perhaps less famous but more similar.
- I really like Messiaen so I let it play me Messiaen "similar artists radio". The result was mostly a stream of tacky sentimental classical music in straightforward major keys, nothing like the sound of Messiaen.
- Similarly, I went starting from the 20th-century composer Xenakis, and it basically played me an arbitrary set of 20th-century composers, none of the pieces sounding like Xenakis' stuff at all. I can think of rock bands whose stuff sounds more like Xenakis (e.g. some of Sonic Youth's stuff has headed that way), but that sort of interesting connection doesn't seem to pop out of last.fm's database.
- For some reason it also keeps recommending me various weedy folk things. Folk can be good, I'm not slagging off folk music in general, but I'm a bit puzzled because it's recommending me rather delicate-yet-straightforward folk music which isn't really my thing. I don't know what it is about my listening history that leads it to recommend that kind of thing. I've listened to a bit of Belle and Sebastian, but beyond that...
On the other hand, it's definitely popped out some stuff which I didn't know before, and I really like:
- The Blood Brothers - this is the one brill recommendation that I think was just based on other things I've listened to. It played me their track "Devastator" which is absolutely wicked.
- Giya Kancheli - It played me one of his symphonies during the Xenakis-based random walk of 20th-century-classical. His stuff is nothing like Xenakis. His symphony #4 was amazing though, definitely going to buy that.
- You Slut! - I already had one track by You Slut! which I really like; so it played me another one. Not the most magically-inspired recommendation then, but it's good that it reminded me about them - they're wicked, nice mixture of math-rock and heavy.
- Animal Collective - not a recommendation, I just listened to a stream of a friend's tracks and it contained some great stuff by Animal Collective.
- I listened to some side-projects connected to cLOUDDEAD people, such as Doseone, and their stuff is cool so I bought a couple of those things.
So, so far, I'm finding new things but very little of it is due to the machine-learning power of their automatic recommendation system, and (as someone who studies this kind of thing) that's a bit disappointing. For example, I would have expected it to recommend me some stuff which I already like, but which it doesn't yet know that I like. That would be a good confirmation that it's doing something clever. Hasn't happened yet. It's still early days though.
Its similarity measurement is largely behaving like "people who bought X also bought Y", which although useful is not as clever as I was hoping. To start recommending interesting things, it needs to be able to bust out of the coincidental connections caused by certain groups of people happening to have been exposed to certain things (e.g. Oasis and Blur, musically there's no strong connection). It kind of needs to have some kind of reverse of the record-shop-style classification system: for example, last.fm enforces no division between classical and pop/rock, but there's a division due to external factors which kind of leak into the data, so maybe it would be clever if it could balance those out somehow.
Despite that criticism, it's a really nice service, a good way to wander around looking for new music; and it's "non-evil" in that it licences its data in a fairly friendly way (creative commons) to allow you to use its data feeds for other purposes. I wrote a little script to archive its feed of what I listen to, so maybe in a while I'll have a handy little database of music I can do things with.