Actually, I Hate That Song

The attempt to build a music recommender that doesn’t suck.

Music recommendation sites like and Pandora, once consigned to the discussions of a few offbeat tech bloggers, are now heralded as the key to halting the downward slide of record sales. And new sites such as Audiobaba and MyStrands have joined the battle for the lion’s share of the online music market. Even so, most music recommenders are battling a creeping paralysis, and are regarded as little more than parlor tricks. What will it take for these platforms to be embraced as the real thing?

The stakes are high. A new, successful platform, parked on the front of iTunes, for instance, could guide millions of consumers down the Long Tail, invigorating huge swathes of the niche market on the way.

The problem with music recommenders is partly rooted in a conceptual split between analytic recommenders, such as Pandora, and collaborative-filtering sites, such as Pandora uses the data of the Music Genome Project to allow users to find music based on its technical values alone—the tempo, for instance, or the backbeat. When Martin Edlund profiled the efforts of the MGP in Slate last year, he observed that the system would help users eschew the superficial components of an artist’s appeal, i.e. “how the lead singer looks in tight jeans,” and concentrate on “precisely defined musical characteristics.” While the idea that we might democratize music by decontextualizing it is certainly noble, most of us don’t listen to music in a bubble. When I hear a song for the first time, I’m weighing the technical ability of the artist, the material’s emotional relevance, and how the music fits into a larger cultural and political climate.

Enter collaborative filters like, GarageBand, and MyStrands, which extrapolate correlations between like-minded listeners to create recommendations. But, as Pitchfork’s Chris Dahlen has pointed out, collaborative filters function like giant magnets: No matter how much you “tweak your recommendations to focus on the most obscure bands,” the results are still going to be governed by groupthink. For instance, if 10 users have just one song in common—a likable hit like, say, Gnarls Barkley’s “Crazy”—this song will keep bubbling to the top of the lists that govern what the radio will play next. A recommender should introduce you to music that you don’tknow. continually yanks users to a well-worn middle ground.

The first solution that comes to mind is a hybrid engine—a platform that could analyze your individual preferences, compare them to the music you don’t already know, and weigh external factors such as blog hype.The more likely scenario is that we, as listeners, will have to adjust our understanding of how our tastes will be altered by recommenders. Last month, I visited Paul Lamere, who researches the science of music recommendation for Sun Microsystems and has built a prototype recommender that provides a glimpse at how totally computerized taste might operate.

On his laptop, Lamere brought up a black screen covered in hundreds of colored dots, each representing a different music genre. This is his library. He gives me an example: Let’s say I’m driving from work to home. On Lamere’s engine, you could plug in the time of your journey and a musical starting point—say, metal. Then you’d plug in a musical ending point as you neared your home—say, electronica. Lamere presses a button, and an automatically generated line zips through 20 songs before winding up at the target.

A similar engine could work inside your iTunes, and it could also work inside a giant library like the iTunes music store. It might also combine the best aspects of collaborative filtering and analytics by allowing specially flagged content to become more likely to be “discovered.” Imagine Lamere’s screen, with its hundreds of dots. Now imagine that the dots, instead of being of uniform size and shape, are of a dozen different shapes and sizes—the music that’s being talked about the most, for instance, is the biggest. Same with the songs that match my penchant for low-fi indie pop. The algorithms that drive the recommendations have been weighted, depending on my personal preferences or listening profile.

So, let’s say a newfangled recommender takes off. What are the implications? First, the large-scale “depackaging” already initiated by the advent of the iPod would be expedited. Bands would spend less time on sequencing since a computer would be combing songs for individual characteristics that have nothing to do with the rest of the record. Musical content would “float” in libraries, waiting to be scooped up by a recommender.

We’d also experience a widening of the industry and more opportunity for indie artists to be heard. Although collaborative filtering would still build and sustain hype, a good recommender would also continuously point to more obscure artists. I currently have 6,200 songs on my iTunes, and I regularly listen to 500. The iTunes store has millions of songs, and if I spent all my spare time clicking through the store, I’d never get through half. But a very good recommender, plugged into my library, would adjust for location, mood, and for my listening profile. It would draw hundreds of lines through my music, generating playlists of songs I’d never otherwise listen to. Only one question would remain: Would my taste in music still be my own?