Researchers all over the world have started using Last.fm tags (which are available through our open API) in their studies. So the next time you tag something cheese on toast or woopwoop you might actually cause some researchers somewhere out there sleepless nights while they are trying to understand what kind of music those tags define ;-)
Here are some research papers that were presented at the International Conference on Music Information Retrieval in Vienna last week and that used Last.fm tags:
- Eck, Bertin-Mahieux, & Lamere (USA), Autotagging music using supervised machine learning: In this paper the authors describe their work on algorithms that try to tag music like humans. I saw a demo and the results are impressive (but I doubt they’ll ever get woopwoop and related tags right). To learn more about their work check out Lamere’s blog.
- Geleijnse, Schedl, & Knees (Netherlands & Austria), The quest for ground truth in musical artist tagging in the social web era: The authors use tags to compute the similarity of artists and classify them into genres. Geleijnse also made some excellent points on why tags used by a community of listeners are so interesting (compared to categories invented by experts who might not even like the music they are talking about).
- Hu & Downie (USA), Exploring mood metadata: Relationships with genre, artist and usage metadata: The authors use Last.fm tags to gain more insight into how mood tags are linked to other tags (e.g. “music for exercising”) and artists.
- Hu, Mert, & Downie (USA), Creating a simplified music mood classification ground-truth set: The authors used Last.fm tags to set up test sets which are used to improve algorithms that can classify moods.
- Levy & Sandler (UK), A semantic space for music derived from social tags: The authors use tags to compute music similarity and investigate important dimensions of similarity. It’s nice to see our neighbours in East London doing such interesting work :-)
Comments
Paul
1 October, 21:38
We may not be able to get woopwoop right, but we certainly will be able to tell whether or not an artist should be labeled as brutal death metal . And thanks for making all of the wonderful tag data available..
Erik Frey
2 October, 00:10
Actually, if it’s woopwoop of the “can I get a” variety, Paul’s audio analysis might work very well!
James Holloway
2 October, 21:12
I’m one of the more prolific woopwoop taggers. A few of us began woopwoop a while back as a fun thing. It began as a mailing list where a few of us could share tunes that we thought would make others sit up and think “this is amazing — i must tell everyone I know about this song right now!”. We were already heavy Last.fm users, and it just sort of dawned one day that applying a unique Last.fm tag was a much easier and more interesting way to share music amongst the WoopWoop group.
This post raises interesting questions about folksonomies. I think research analysing how Last.fm tags correspond to music styles is incredibly interesting, but a starting assumption that tags must correspond with a type of music would be false (not that any one is doing that, necessarily). I think an assumption that music with a certain tag has to have any unifying characteristic is equally false, for that matter. Part of the fun of tagging is the quirks and aberrations that crop up, and either disappear as quickly as they appeared, hang around for a bit like a drunken hobo, or fully catch on. I’m immediately grabbed by cheese on toast because I don’t initially get it. When I look at the tag’s listing, I see Gloria Gaynor’s ‘I Will Survive’ and understand: okay, this is asupercheese tag. If I were looking to create one myself I might have used stilton, but this is cool and I’ll probably use it hereon. Similarly, I’m more inclined to listen to a brutal death metal tag than a hard rock tag, paris hilton and barney spikes included. Using the obvious/popular tags is very useful too, of course.
I don’t necessarily accept the idea that vandalism exists in a folksonomy. Finding Paris Hilton in the brutal death metal tag might be incongruous or downright annoying for brutal death metal fans, but there’s probably a large proportion (and much greater number) of hip hop fans annoyed by the presence of 50 Cent in thehiphop tag top ten. The difference in the Paris Hilton case is not the number of people for which this is a problem, but how obvious it is to third parties that an annoyance exists. Annoyances infolksonomies occur all the time.
With the woopwoop tag, we’ve tried to create a common thread between songs; it’s just not genre- or style-specific, it’s about that indefinable quality in a song that makes one person want to tell everyone they know how good that track is. Hopefully, as people tag new tracks and double up on those already tagged, this will make things more useful. Of course, you can’t prescribe tag usage so it will be interesting to see ifwoopwoop adapts in any way. The idea that there is a unifying characteristic to all the woopwoop tracks that we hadn’t realised is really appealing, as would any research that can analyse the cheese on toast tag and spit out a recipe for the ultimate musical wedge of stilton.
By the way, the name we chose was shamelessly stolen from a track by The Chap, for various reasons.
Paul
2 October, 23:22
James: Good comments – I’m a big tagging fan as well, and agree – we don’t actually start with the assumption that tags have a direct relation to the music … in fact the work from UCSD was looking exactly at that issue, trying to find out which tags had some correlation with audio and which ones didn’t. It makes no sense to try to learn how to predict the ‘lazy eye’ tag from audio if there’s no correlation between the tag and audio.
I do have to disagree a bit about vandalism though … some of the tags that are applied are like graffiti … it is art to some, but to most it’s an annoyance. That doesn’t mean that we should try to prevent people from applying those tags. We can learn much just by noticing which artists are subject to such attacks. I think, however, that it is important to try to filter the tags when they are used. From last.fm’s point of view this is very important. If they are trying to grow their listener base, they are going to have to attract a wider array of music fans, this will eventually include (as hard to believe as this may seem) some Paris Hilton fans. If the Hilton fans are presented with songs by Cannibal Corpse or Cryptocracy every other song, they will quickly decide that last.fm is not for them.
pilgrim
3 October, 14:47
Paul, your point about Paris Hilton fans being presented with songs by Cannibal Corpse etc is well taken. A general solution to this problem would be to allow users to include/exclude tags from consideration when compiling recommendations, no?
I must admit I personally don’t tag music in last.fm, and I don’t often consult last.fm for recommendations; for one reason or another, the usefulness of the system breaks down a bit for users who depart from the norm in the ways that I do (I’m probably waaay older than your typical user, and I have truly stupid quantities of music already, across a very broad range of genres, and ranging through the entire history of recorded sound)
I’ve implemented a kind of tagging system in iTunes itself – the details of which would be too involved to describe here. Possibly I’ll get around to posting about it in my blog.
Norman
3 October, 22:05
I agree with Paul: folksonomies are great as long as a large fan base is not hijacked by a small but organized group of haters. This does not mean we censor, but simply that our filtering algorithm works (mainly) by the principle: “you have to eat from the dish you spit into”. ;)
@pilgrim: have you been checking our recommendation/radio lately? We’ve been putting a lot of effort to please non-mainstream users (as one of them).
Endre
4 October, 15:48
Some months ago, when I signed up after Pandora suddenly told me to fuck off (I don’t live in the World .. eh, I mean, USA), the standalone player had this AMAZING tag-cloud thing. This stuff would display other “similar” (related, one would believe) tags when tuning into a tag.
However, this is now gone. Why?! I don’t really believe the algorithm is so amazingly hard, really..! :-)
Clouds of both artists and tags, both when “browsing” artists and tags, would be VERY VERY good. I miss it so extremely much. Please add back. (Or, if this is some configuration option I’ve suddenly turned off, then please enlighten me).
Jon
4 October, 21:16
Cool – we started a Last.fm Research group specifically for stuff like this – the use of Last.fm data in music research. It’s exciting to hear of folks actually doing some interesting research and using the data. I hope some researchers find their way over there so that we can discuss potential collaborations and future ideas.
Nat
30 October, 08:09
I’m also looking at using Last.fm data for analysing mood as part of my research into ubiquitous computing.
I’d really like to be able to rate and query music on emotional scales, perhaps using the PAD scales.
http://www.kaaj.com/psych/scales/emotion.html
I think this would build up an extremely valuable dataset.
Comments are closed for this entry.