Brain mining
Nov. 21st, 2004 11:20 am![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
mmcirvinAs I've mentioned before, for a while now I've been using a collection of iTunes/iPod Smart Playlists whose criteria generally involve picking out the least-recently-played songs above some play-count threshold, or within some range of play counts, possibly with some other criteria applied. Using them as a sole means of listening to music would, of course, be pointless, since they depend on a preexisting database of play-count and time-stamp information reflecting your personal taste. But they're a great way to surprise yourself with things you haven't listened to lately, but like.
Today I was listening to one of these lists, sorted by play count (and secondarily by time last played, though iTunes doesn't give direct user control over n-ary sort criteria), and the list produced a succession of tunes that were amazingly thematically connected, all brilliant, dark songs about people with screwed-up heads. And I marveled at the coincidental wisdom of its mechanical brain, until I looked at the "last played" fields and realized that the reason they all popped up at once was that I listened to them last in the same half-hour period in mid-October. And the reason they all have similar play counts is also probably that I listen to them when I'm in the same mood.
I think that some or all of them may have also once appeared near each other in a long-vanished version of one of my manual playlists; though, had I not actually done much listening to that part of the playlist, the resulting play-count correlation would not have happened.
So I conclude that this automatic playlist strategy actually works remarkably well at dredging up information from my own behavior.
But it's still not as elaborate as it could be. The next step for iTunes (or somebody else, if it hasn't already been done) is to start explicitly collecting time correlation data. Of course there are plug-ins that upload your play counts to central servers so that other people can get "people who liked this also liked..." information. But I'm thinking that it should be more time-based, so that you actually have persistent "when you liked this, you also liked..." information; since my effective musical tastes change radically depending on mood. Then you could generate Markov-chain playlists. Actually, I wouldn't be surprised if this made taste-correlation databases accumulated across many people much more useful.
I dunno... does Audioscrobbler or something already do that?
Today I was listening to one of these lists, sorted by play count (and secondarily by time last played, though iTunes doesn't give direct user control over n-ary sort criteria), and the list produced a succession of tunes that were amazingly thematically connected, all brilliant, dark songs about people with screwed-up heads. And I marveled at the coincidental wisdom of its mechanical brain, until I looked at the "last played" fields and realized that the reason they all popped up at once was that I listened to them last in the same half-hour period in mid-October. And the reason they all have similar play counts is also probably that I listen to them when I'm in the same mood.
I think that some or all of them may have also once appeared near each other in a long-vanished version of one of my manual playlists; though, had I not actually done much listening to that part of the playlist, the resulting play-count correlation would not have happened.
So I conclude that this automatic playlist strategy actually works remarkably well at dredging up information from my own behavior.
But it's still not as elaborate as it could be. The next step for iTunes (or somebody else, if it hasn't already been done) is to start explicitly collecting time correlation data. Of course there are plug-ins that upload your play counts to central servers so that other people can get "people who liked this also liked..." information. But I'm thinking that it should be more time-based, so that you actually have persistent "when you liked this, you also liked..." information; since my effective musical tastes change radically depending on mood. Then you could generate Markov-chain playlists. Actually, I wouldn't be surprised if this made taste-correlation databases accumulated across many people much more useful.
I dunno... does Audioscrobbler or something already do that?