Among the algorithms that run our online lives, Pandora's Music Genome Project may not be as critical as Google's search equations, but the math behind the self-programming Web-radio service seems just as opaque.
Some of that mystery is by design, but Pandora's been a little more public about how it gathers and grades the more than one million tracks in its collection. And over a long briefing at its Oakland offices and subsequent follow-ups over e-mail, it told me a few more details.
First, humans grade the songs, not computers–and it's surprising how much effort is involved. Pandora's 25 or so music analysts have to assign either one-to-five rankings or more quantitative measures (say, beats per minute) for as many as 450 "genome units" per song.
For example, vocals get graded on terms like "Smooth or Silky" (Tom Waits' "Come On Up to the House" earned a totally unsurprising 1), "Delivery Spoken-to-Sung" and "Child or Child-like." Music can be broken down by metrics such as "Melodicism Lo-to-Hi" (how easily could you play this back?) and "Melodic Articulation Clean-to-Dirty" (how precisely does the melody hit the beat?).
Each of Pandora's analysts–many hired from the Bay Area scene via word-of-mouth–can handle four songs an hour, for about 10,000 songs graded a month.
Is it possible to listen to music like a normal human being after you leave the office? "It's something I can turn on or off," said analyst Steve Hogan during my briefing in May. "It hasn't interfered with my ability to enjoy a piece of music."
(In his spare time, he plays the organ at San Francisco Giants home games.)
Secomd, they don't care about your location, or if you buy music through the service.
I'd like to be able to request songs that come from artists near me–locavore listening!–but Pandora reflects its origins. It's a service that predates location-aware smartphones.
"One of the themes that runs through [our listener feedback] is the ability to localize the music selections," said founder Tim Westergren. He allowed that the bigger potential might be at the other end of the system–letting musicians see where they're finding an audience. "Boy, what if you could allow a artist to log into Pandora and see a heat map of where their fans are and plot out a tour?"
Musicians can make money more directly from an engaged Pandora listening by selling tracks through the service's click-through links. You might think that this transaction–the most direct fan support possible through the app–would weigh heavily in its calculations, but it's not factored in either.
Apple's iTunes Radio makes for an interesting comparison, since it can start to customize music stations not based on one artist but on the content of a user's entire iTunes library.
Finally, negative feedback is given more weight than positive feedback.
Years ago, Pandora let its listeners' use of the thumbs-up button drive what what they'd hear. But as more of its usage has shifted to mobile devices, it's found that people "thumbs-up" less but hit the thumbs-down and skip buttons as much as ever.
"We stopped chasing thumbs-up and started chasing total listening hours," said Eric Bieschke, head of Pandora's playlist engineering team. For a song to serve that goal, it had to rank highly in four areas–some of which seem to contradict each other.
Prediction accuracy, in the sense of matching the right music to the right listener, seems to clash with the virtue of " surprisal" (Bieschke's term). And variety, taken to a sufficient degree, conflicts with relevance.
And that's where Pandora lets me down most often. Heard over hours at a stretch–for instance, when I'm testing a smartphone's battery life–its talent pool can look a little shallow. Not just the same artists but the same albums crop up repeatedly; for instance, the Pandora station I modeled after my friend Eric Brace's band Last Train Home seems guaranteed to serve up Ryan Adams and the Cardinals' "Cold Roses" within the first two hours.
In a 3,172-word post in 2010, University of Virginia music researcher Jason Kirby, pursuing a Ph.D. in music, poked at a similar hole in Pandora's output. "I find homogeneity of songs' tempo an issue," he wrote. "The Genome's platform, as it currently works, seems unable to deliver the ebbs and flows in tempo and musical texture which I enjoy in a good mix tape or college radio show."
Alas, mixtapes are dead, and college radio can be heard nowhere near me.