Subscribe
A Crisis of Faith
February 7, 2007 12:28 am
I wrote about Pandora earlier — a music recommendation and web radio service that hooks you up with music that shares similar qualities to the band and song you seed your “station” with. In the comments, Dragon wrote about last.fm. I hadn’t actually used last.fm at the time, so I filed it away in my brain under “totally web 2.0 services I should try sometime.”
I’ve been playing a lot of WoW (I need to write about why I quit EQ2, anyway), and I’ve been taking a lot of painkillers for my tendinitis, so I’ve been in the mood for new music. Today, I tried last.fm.
last.fm is a music recommendation and web radio service that hooks you up with music similar to other music similar to your seed, the same as Pandora. But while Pandora looks at the actual qualities of the music you’ve entered, last.fm data mines the hell out of users’ music collections. It’s nature — the actual qualities of your chosen band or song — vs. nurture — what other people who like your band or song like.
In my experience, Pandora kicks the shit out of last.fm. Actual human effort is kicking the shit out of data mining.
My “user story:” I’ve been listening to the same industrial bands for the last ten years. I love me some KMFDM, but I love me some Pig more — Pig is much like KMFDM, given that they share a member, but Pig is heavier. I like guitars, drum machines, and samples. Pig is, absolutely, my favorite band, and has been for a very long time. For most of that time, I’ve assumed that I like industrial music, but every time I’ve looked for new industrial music, I haven’t liked it. It’s been about synthesizers. I just assumed that the current state of industrial has passed me by. I assumed that I was getting old, and people just weren’t making it like they used to.
A few months ago, I loaded up Pandora and entered “Pig.” Its human-generated music recommendation engine recommended industrial metal bands, something different than what I’d been manually looking for.
My mind was blown. There were other bands out there that I liked — in particular, the Kovenant and the Deathstars. I mean, oh my god, I could have been listening to honest-to-god new music, and I would have liked it. Oh my god, I’m not just getting old!
Since Dragon mentioned last.fm, and since I was in the mood for new music while I slogged through quests on heavy narcotics, I tried it out today. I have my old standby, my old favorite band to try out, and I have these new bands as well. last.fm data mines peoples’ music collections. And even though, as I pointed out in my last post, Pandora’s community recommendations looked a little suspect, I figured that since last.fm’s business model depended on data mining, all would be well.
This wasn’t true.
I started my last.fm experience by entering the Kovenant. I got heavy black metal shit.
Well, okay, the Kovenant started out as a real black metal band, and I only really liked their last album (really, I think SETI is an incredible classic of industrial metal). last.fm, unlike pandora, gives me no way of refining what I like about the band I entered.
Okay, I’ll try something else. I tried the the Deathstars, another kind-of-but-not-really black metal band.
I get more … real black metal.
Okay, I’ll get back to basics and enter Pig.
I get EBM, a subgenre of industrial that I really can’t stand. Well, shit.
The problem is that last.fm is a data mining service without constraints — anybody who subscribes to the service can submit all kinds of data, and as I suspected with glancing at Pandora’s community pages, people suck. People who like Pig don’t share my other tastes. Web 2.0 hipsters who like Deathstars and the Kovenant don’t share my tastes. There are no constraints on one’s musical tastes.
last.fm is data mining wholly unconstrained choices, where people’s music collections are wholly free to share space with bands I like and bands I don’t.
Pandora isn’t data mining at all, and their “music experts” are contraining users’ music stations amongst choices they’ve personally evaluated.
I don’t think I’ll be listening to last.fm again. People are idiots.
Edited in the morning, 02/07. What was I thinking?
Dragon wrote:
Woah! Almost wish I’d never mentioned it now.
Still, it’s a good post so it’s not all bad that I mentioned it. I have to confess that I don’t remember the last time I used last.fm though.
Funnily enough, the only way I got to a channel I liked in Pandora was by entering White Zombie as the initial starting choice. Any of the other bands I like threw up loads of stuff I didn’t but the White Zombie channel played more of what I was into and introduced me to loads of new stuff as well.
Will be interested in hearing about the reasons you gave up on EQ2. I know what mine were!
Posted on 07-Feb-07 at 5:29 am | Permalink
Darius K. wrote:
I use last.fm not for recommendations, but simply for the access I get to data it collects on my own listening habits. For that, it’s pretty great.
As a side note, I always wanted a music player that did that kind of collection locally. I’d like my music library and listening habits to be literally stored in a local SQL DB–then I could do things like
SELECT Songs.Artist, COUNT(*) FROM Songs, Plays
WHERE Songs.Id = Plays.SongID
AND Rating IS NOT NULL
AND Plays.Number > 10
GROUP BY Artist
and create my own damn analysis of my listening habits. That would be geek heaven.
Posted on 07-Feb-07 at 7:10 am | Permalink
Stephen C wrote:
Last.fm sounds like it does something similar to MovieLens (http://movielens.umn.edu/). With MovieLens, after rating about 100 movies all the suggested ratings for other movies were within 10% of my actual rating, and they just kept getting closer. So the theory seems to be sound in certain situations.
Personally, I had just the opposite reaction to Pandora; it would give me a huge amount of stuff I couldn’t stand for every song that I liked, and I wished it used some sort of data mining to supplement its conclusions. Could the problem with last.fm involve a poor matching algorithm or an insufficient sample size for your type of music, rather than a flaw in the idea of using data mining to suggest music?
Posted on 07-Feb-07 at 8:29 am | Permalink
Aaron wrote:
I’ve been using Pandora on and off for months, and I’ve had mixed results.
I can’t help but think that you’d need 6 months or more of categorizing individual song elements, like they do, before developing a fully nuanced and steady collection of categories. For example: I might begin by observing “dark melodies” present in Testament songs. But once I got to considering Pantera and Alice in Chains, and realized they both have songs that could be said to contain “dark melodies”, I’d want to find a more precise terminology. It would take a long process of trial and error to hone my terminology and apply it consistently.
Perhaps I’m wrong, but I get the impression that there was not enough of a testing period for Pandora, so a lot of their initial, flawed classifications are mixed in with the better ones. And many of their classifications are not that precise (I think “dark tonality” is one they use).
Still, I’ve found a couple bands and more than a few songs I really like through Pandora. I’m a fan. My cousin, perhaps because he’s less picky than me, was an even bigger fan when I introduced him to it.
Posted on 07-Feb-07 at 11:46 am | Permalink
Sara Jensen wrote:
Ooh, I think I got it. Maybe the thing that Pandora gets right is the ability to specify multiple seeds — I create a station that’s seeded off the Kovenant (a real live metal band) and Pig (a real live industrial band) and I get results that combine the best of both, exactly what I’m looking for.
At last.fm, I can only enter one seed, and I get stuff that’s too far in either direction. What I would rather see is matches based on the habits of people who liked both the Kovenant and Pig.
Posted on 07-Feb-07 at 12:22 pm | Permalink
Aaron wrote:
Really, I suppose it would be nice if I was given the option to select which characteristics of a song or band are the ones I want replicated in another song or band. So, for example, I could tell the program to find a band with guitarlines similar to Pantera’s, but without asking for bands with vocal parts similar to Pantera’s.
Posted on 07-Feb-07 at 1:08 pm | Permalink
Dragon wrote:
“Maybe the thing that Pandora gets right is the ability to specify multiple seeds ”
yes, yes, yes.
I went back to it today and started “training” a new stations. What I’d love to do is put a couple of seeds in or relate my bookmarked songs to particular stations and not others. (I notice in your profile that KMFDM is both the most thumbed up and thumbed down - conflicting results or what)
Posted on 07-Feb-07 at 2:57 pm | Permalink
Rahul wrote:
I listen to the Kovenant as well as use last.fm, but I think your conclusions of last.fm from what frankly was a pretty short-lived experiment are a bit out-there.
I think it’s more that you’re trying to use last.fm in the same way that you use Pandora, which obviously won’t work since last.fm operates on the premise of social interaction, whereas Pandora approaches music technically. So what you’re looking for when you run a search on last.fm for “similar to the Kovenant” isn’t so much music that sounds similar to the Kovenant, but “people who we think have a similar music taste to you based on your search parameter of The Kovenant”.
Obviously that also won’t get you very meaningful results, because hey, plenty of people may happen to like this same band as you do but hate everything else you listen to. So for last.fm, you really need to broaden your search parameters in the same way that you’d not just accept single by-chance “omg we like the same band” from a random person (such as me!).
Last.fm, then, is a great way to meet people, not so much find new music. That’s the key differentiator between it and Pandora. Pandora is focused entirely on getting you directly to the nearest-sounding band on the list. Last.fm is going to get you to the nearest person. Last.fm is a web 2.0 style social software experiment, whereas Pandora is a music comparison tool.
Still, though, great to randomly find another appreciator of the Kovenant. You may care to know that they’ve supposedly been working on their next album for a while and those of us who keep track expect it to appear next year. Strangely, from my perspective of the band, their greatest work was in fact not SETI, but Nexus Polaris, a symphonic black metal classic from the late nineties, and possibly the album following that, Animatronic, which was more towards your style of music with its industrial experimentation. Perhaps Aria Galactica will marry our two tastes, and then we’ll have another thing in common. Maybe last.fm will pick up on it
Posted on 08-Feb-07 at 3:34 am | Permalink
Sara Jensen wrote:
You’re right, Rahul. But I think both services present themselves as solutions to the problem I’m trying to solve — I’m tired of listening to my old mp3s while I play WoW. I’d like to not only listen to different stuff, but I’d like to be introduced to new stuff that I’ll like as much as the old mp3s.
Is there a way to refine your findings on last.fm or to create a station with multiple seeds? I see that I can “ban” stuff.
Posted on 08-Feb-07 at 12:32 pm | Permalink
Rahul wrote:
If you download the last.fm software, you can create dynamic radio stations using a variety of options — tags, artists, based on what your neighbours are listening to, based on your own tastes, based on your friends’, etc. It’s pretty versatile. You can ban/love stuff to nudge it around a bit.
Make it a habit to have last.fm open and tracking your listening habits whenever you listen to music on your computer. That way it’ll slowly build up a profile of you and be able to recommend more/less relevant things since it’ll know more about you. It’ll also generate more relevant neighbours, etc.
And hey, shouldn’t you be listening to the in-game music while playing WoW? The Burning Crusade is completely orchestrated, you know.
Posted on 09-Feb-07 at 6:07 am | Permalink
Joshua Gerrish wrote:
Yeah, one of the things to keep in mind with last.fm is that it’s best at recommendations when it has more than one data point, and it collects those data points from your listening history, not your browsing history. It’s a personalized collaborative filtering or recommender system.
Posted on 02-Mar-07 at 1:20 am | Permalink
Anonymous Coward wrote:
You’ve clearly misunderstood how you’re supposed to use last.fm
The idea is that you log your music listening habits over a period of time and when you have say 1.000 tracks logged last.fm is able to make a good match of your composite taste against the aggregate, composite taste of other people.
Entering one band name and expecting to be able to infer your own taste on the basis of that is a pretty misguided understanding of how datamining works.
So please do your homework properly before posting a negative review of last.fm or at least make your assumptions of how the service is supposed to operate more explicit.
Posted on 07-May-07 at 3:02 am | Permalink