the perceptron: astute music recommendations
August 13, 2008 7:31 AM   Subscribe

the perceptron: astute music recommendations
Enter a band or artist you like, get back five recommendations with descriptions and sample tracks. The site suggests bands based upon real people's music tastes. Its sources include recommendations made by other humans, artists on the same label, artists who have played shows together, artists who admire each other.
posted by maryrosecook (22 comments total) 1 user marked this as a favorite

I've been playing with this for about twenty minutes, and it's surprisingly good at doing something that Pandora does not: recognize links that do not involve sounding exactly the same.
I suppose that 'sounding the same' is actually a sort of weak metric for music most of the time, and by using things like having played together and being labelmates, it's not like you're going to type in Iron and Wine and have two of your hits be death-metal bands.

Well done. This has been bookmarked, and will be thrown out the next time someone posts an AskMe for music recommendations.
posted by The Esteemed Doctor Bunsen Honeydew at 12:31 PM on August 13, 2008

Thanks for much for the feedback and kind words. Yes, this is exactly what I'm shooting for. I have often been surprised when bands I like name the bands they like - they often choose artists in a completely different genre. This led me to thinking about how weirdly and diverse many people's music taste is. So it seeemed that the best way to capture this was to capture specific people's music tastes: that of someone who organises gigs, or runs a label, or is in a band you like.

I'm really keen to hear about ideas for improvement.
posted by maryrosecook at 12:48 PM on August 13, 2008

I played around, and it worked great for even obscure bands. I had some inconsistencies when I typed "Black Keys" vs. "the Black Keys" though.

What are your data sources, and how do you associate? It would be a cool project to associate artists based on if they were mentioned in the same blog post... idea!
posted by ajcronk at 2:13 PM on August 13, 2008

Yep, the search system is currently *very* naive. On the list is full-text search, misspelling suggestions and optional definite and indefinite articles.

Your question is really interesting. I basically try and sample real people's music tastes. For example, I scrape Myspace and work out which bands are friends with each other and, thus, form links between bands who like each other's music. I scrape The Hype Machine and link bands who are mentioned on the same mp3 blog. I also get in some recommendations from a couple of sites that have human recommenders on them. Next on the list is scraping concert listings to link bands who are playing on the same bills.

However, there is an extra subtlety. Some of these sources of information are going to be great, and others are going to suck. Thus, the relative weights given to sources is adjusted based upon which links are "loved" by users.

I've also got some ideas about skewing recommendations towards my idea of a user's taste in music.
posted by maryrosecook at 2:49 PM on August 13, 2008

Excellent name and interface.

As for the recommendations, it would be nice to get the name of the sample song and the name of the album on which it appears. offerent 2 or 3 songs per artist would give them more of a chance... as it is, we have to decide whether to visit their website after a single song.

Your recommendations are pretty hit-and-miss: they were pretty good when I searched for Beirut, but terrible when I searched for Lila Downs. Searching for Lhasa de Sela seemed to bring in a smattering of various music by French/Canadian artists, but most weren't terribly related.

I can't tell whether this problem will be fixed as more people click the "love" links. The thing is, they're going to think that "love" means "I enjoyed this" and not "this is a good match." There isn't any way to downmod recommendations that aren't any good.
posted by Grimp0teuthis at 7:24 PM on August 13, 2008

I really wanted to give this a whirl, but I got a "502 Bad Gateway". What gives?
posted by soundofsuburbia at 5:17 AM on August 14, 2008

Site went down when I was asleep. Very sorry. It's back up now.
posted by maryrosecook at 5:29 AM on August 14, 2008

Hooray! I do like this! Even though there were zero results for my girlfriend's band! But "Sun City Girls" returned "Thinking Fellers Union Local 282" and thats spiffy enough for me!
posted by soundofsuburbia at 5:34 AM on August 14, 2008

Great! So glad you like it. If your girl's band is on Myspace, we'll get them indexed soon. If you tell me their name, I'll put them higher up the list.
posted by maryrosecook at 5:39 AM on August 14, 2008

That's so nice of you! :)

Here's her MySpace. She's pretty unknown (at least at the moment, but that will change or I will eat my hat), but she can be found at The Hype Machine and at MySpace (of course), so I figured I should see if she'd made it to The Perceptron!
posted by soundofsuburbia at 6:01 AM on August 14, 2008

Here you go!
posted by maryrosecook at 10:22 AM on August 14, 2008

Things get weirder the further into the past you go: "Captain Beefheart" recommends Chic? I guess, sure, OK, in that I like both of those bands… But it doesn't mention Frank Zappa? "Miles Davis" recommends John Denver and Donovan? "Elvis Presley" recommends Os Mutantes (and vice versa)?
posted by klangklangston at 1:49 PM on August 14, 2008

Yes, the connections can be weird, but that is kind of the point. Some of the recommendations will stay within the genre of the band you typed in, but many will not. This reflects the data sources, which are all based upon real people's music tastes.

However, as the site gets more popular, the recommendation logic will tune itself based on the feedback it gets from users. As well as the 'love' link, I am considering adding a 'hate' link.
posted by maryrosecook at 2:49 PM on August 14, 2008

It's not just genre: There are simply better matches for Captain Beefheart than any of those shown. The reason why Zappa should be listed isn't because Zappa is in the same "genre," (both because he's generally not, and because both Zappa and Beefheart spanned a lot of musical territory), but because Zappa and Beefheart had a long working relationship. That's kind of the explicit linking that I felt was promised. If someone likes Zappa's "Hot Rats," which features Beefheart, they'll probably also like, say, Tony Williams' Lifetime, which shares milieu. Or any number of side projects from the folks who were in The Magic Band.

Likewise, while Miles Davis probably suffers from being over-broad (though I'm not sure how to datamine properly to correct for artists like Davis, who have extreme breadth in their careers), John Denver isn't a weird connection—he's no connection at all.

Two things might help this: First, improving the granularity of artists like Miles Davis to try to get better at quantifying WHY people like them, as the similarities in taste aren't simply a matter of nodal-distance when there are underling qualities that are better predictors. Second, increasing your data pool to include record reviews would really help, as that's a pretty big pool of professional recommendations based on explicit reasoning. Perhaps running a search for whatever band in this hypothetical review pool and noting when other bands that have been searched for are also mentioned in the same review. This, of course, has the disadvantage of requiring a pretty huge index…
posted by klangklangston at 3:18 PM on August 14, 2008

maryrosecook: Sweeeet!
posted by soundofsuburbia at 1:15 AM on August 15, 2008

Your comments are really insightful - thank you.

I think splitting up artists into different parts of their careers and styles they play in is a fascinating idea. I have some ideas around personalising recommendations so that they skew towards certain connections rather than others. For example, if you said you like Converge and had "loved" a bunch of indie stuff (and were thus probably a fan of the You Fail Me album), the site would skew the recommendations towards bands closer to the other bands you had already said you like. This might help with the Miles Davis Problem.

Your second idea is one that I've had in mind for a while, and I'd very excited about it.
posted by maryrosecook at 1:24 AM on August 15, 2008

The Miles Davis Problem would be a good (though pretentious) name for a band.

But yeah, as someone who's worked as a professional music critic, how to make recommendations is something I've thought a lot about, and I've been kind of disappointed by every metric that I've seen so far to automate the process (Pandora, Last.Fm…). They all have some methodological failures that skew things one way or another. Even in personal consultation, there are a couple of problems: The difficulty of quantifying every connection, and the fact that most people who ask for recommendations aren't very clear on what they want or what they like. There's a failure of articulation.

That, and frankly, there's just mountains and mountains of music out there. Which means both that it's hard to know enough to make accurate recommendations, and that recommendations tend to skew upwards toward more popular artists. This is definitely a problem with, and with most automated recommendations—I need to go in the other direction when I want recommendations (I already have all the Beatles I need, Radiohead too), but the degree of confidence falters as you get further out from the node.

Feel free to email me if you want to talk more.
posted by klangklangston at 8:57 AM on August 15, 2008

Great - thanks. Will email for more discussion.
posted by maryrosecook at 3:16 PM on August 15, 2008

This is pretty cool, but it doesn't understand that when I type "beatles" I mean "THE beatles". There are so many "the" bands - do I need to type it every time?
posted by easternblot at 3:54 PM on August 17, 2008

Thank you for your feedback. Really glad you like the site.

You actually don't have to type the "the" for The Beatles or any other band with "the" in their name. The reason it wasn't working was an error that just affected the recommendations for The Beatles. I have now fixed the problem.
posted by maryrosecook at 5:46 AM on August 18, 2008

A long time ago I read an article where the author mentioned that some record company was failing to "market" a new Jazz performer. They just seemed to be unable to get the usual Jazz crowd interested. They went to some sort of music industry consultancy firm (I think this one) and they suggested that the record company should try to sell the artist to Marilyn Manson fans. The record company didn't like the idea but gave it a shot and it was a success.

I can't remember if the article was this one, but it has to be that company. I remember it because the CEO has the same name as one of the guitarrists in Pearl Jam.

But the thing is, some of the basic idea seems the same, to make clusters of "likeness", no matter how disparate the sound of the bands/artists may seem to the ear. There is a whole different angle to their invention, but it might still be interesting to you. Another thing you might want to check out is the CEO's own blog.

Cool thing, the perceptron. Good luck with it and I hope you find the links above interesting.
posted by micayetoca at 8:48 PM on August 23, 2008

Wow - that is so interesting. I'm going to check out those links now. Thanks a lot.
posted by maryrosecook at 3:20 AM on August 25, 2008

« Older A Winner Is Me!...   |   1979: 29 Dorks... Newer »

You are not currently logged in. Log in or create a new account to post comments.