Predicting bands a user is going to like isn’t easy. But surely Spotify, Pandora and iTunes can do better than this.
I’m a freak for new music. Always have been. In a given day I’m usually listening to whatever cool stuff I have discovered recently, backtracking and catching up on bands I haven’t listened to lately, and trying to find new artists to fall in love with and suggest to my friends.
Finding new music is a different challenge than it used to be. Once upon a time you could turn on the radio and hear the latest and greatest. It’s been a long time since that worked, though – now radio is the last place you look for cool tuneage.
Another strategy that used to work: record stores. When I was in grad school at the U of Colorado the awesomest indie music store on the planet, Albums on the Hill, was close at hand. They got to know me and were consistently able to turn me on to bands I was going to enjoy. But then the Internet killed the music store.
Word of mouth has always been important, of course – I have friends who point me to new things, like recently when in the course of a couple days Carole McNall found St. Paul and the Broken Bones and Rich Flierl introduced me to Phantogram. Bob Lefsetz’s e-mail letter has a strong track record of mentioning worthy acts – that’s how I discovered Jets Overhead some years back.
Then there’s Facebook. My FB friends are always posting videos, and sometimes I find a band I like.
Another rich source of discovery is – or at least ought to be – the online music service: Spotify, Pandora, iTunes Radio, eMusic, etc. I say “ought to be” because these places have some Big Ass Data to draw on. When you have catalogs that large and can track the behavior of millions of users, you can theoretically get pretty damned good at predicting what someone might like based on their known preferences and those of others with similar tastes.
I routinely test this theory out, and have set up channels at a few sites for this purpose. I’ll seed a playlist with several artists that share a sound (at least, they do in my head) and hit play. You also have an ongoing passive attempt at getting my attention on Spotify. Their default page is all about discover: “You recently listened to X. Check out Y.” “You listened to A. You might like this song.” Recommended for you: Z.” And so on. Their algorithms are working round the clock attempting to figure out who I haven’t heard yet that I might like. Which is great – I love that they’re trying to do this.
I just wish they were better at it.
A couple things tend to happen. First off, I get recommendations for bands that are nothing like what I’m after. Second, I’ll get recs for bands that are sort of in the right ballpark, but – and I hate to disparage – they just aren’t very good. I’ll sample them – and this has happened at least a couple hundred times by now – and the first track will be decent enough that I think maybe they deserve a listen. So I’ll let it play and go back to work. A couple minutes later the second and third and fourth tracks get so annoying I have to interrupt whatever I’m doing to pop back over and kill it. This happened three or four times yesterday alone.
I understand that recommendation algorithms are anything but a precise science. The core data, I’m guessing, is going to revolve around who listens to whom. They’ll start by comparing the groups I listen to with other users, and when they find the people most like me they begin suggesting the bands that are on their lists that aren’t on mine.
For instance, if Bongo237 and OldSkoolStef listen to a lot of the same neo-Soul artists, the system may notice that Bongo loves Ryan Shaw and Stef has apparently never listened to him. So next time she logs in her Discover page might include “Recommended for You: Ryan Shaw.” If Stef starts listening to a lot of Ryan Shaw, then the system’s faith in that connection is reinforced (especially if she gives it a thumbs up when he pops up on her radio station).
Which leaves me wondering about Spotify users in general. Apparently users that have the most in common with me also listen to lots of self-consciously plinky lo-fi hipster indie wanking. My gods, people, get some production values and learn to write a song. “Less is more” doesn’t apply if we’re talking about talent.
Even when these services are recommending good bands, they’re often missing my intent (and suggesting that they’ve never heard any of the bands in question). Check this one:
Ummm. Seriously? How in the hell do you get from The Raveonettes to Belle & Sebastian? Eels is a slightly closer match to B&S (depending on which Eels CD we’re talking about, anyway) but even that’s an odd recommendation. And Eels isn’t a band I’d lump with Raveonettes, either – are they trying triangulate? If so, they’re not doing very well. Assuming you were using these three bands to triangulate, the one in the “middle” would be Eels, and I can fathom no auditory aesthetic that would put B&S “between” the other two.
You make the call. Here’s a popular Raveonettes song.
And now some Belle & Sebastian:
I’m not dogging B&S – they’re quite talented. And I might like them. But it won’t be because I like The Raveonettes. More like despite the fact that I like The Raveonettes.
What I wish these algorithms did a better job of accounting for is the sound of a band or song. If I’m looking for something based on my Raveonettes Jones, I want you to point me to Dum Dum Girls, or Frankie Rose, or The Pains at Being Pure at Heart, or Silversun Pickups or maybe even The Blueflowers or The Lost Patrol. Or maybe go back a few years and suggest someone like Sugar or even one of the Shoegazer bands.
I guess give it a few more years. Musical tastes are complicated, and the perfect algorithm is going to have to account for similar user behavior, of course, and then figure out how to meaningfully understand that sound thing I mention. Getting a tighter grip on ever-evolving genre boundaries will be critical. Then things like understanding music subcultures and politics and figuring out that I don’t ever want to hear from an industry put-up job or anyone who has ever recorded a song by The Matrix and, honestly, if they so much as met Simon Cowell at a party I’m disgusted with them.
I’m trying to be fair here, and I know I’m not the easiest guy to please. I’m okay with the fact that an algorithm is going to miss when trying to guess what I’m in the mood for. But days like yesterday, where it missed 20 times in a row and didn’t hit once, I’m not going to lie – they make me long for a world with smarter technology.