The Million Dollar Programming Prize
Want Advice? Try an Expert
By Philip E. Ross
If you want recommendations and you’d rather not rely exclusively on your customers, you can always do the daring thing and consult actual experts. That’s the idea behind Pandora, a free Internet radio service that employs musicians to rate songs according to a checklist of criteria, such as pace, rhythm, even the voice of the performer.
The Oakland, Calif.based company must be doing something right. In the three years since it opened shop, Pandora Media has registered 22 million listeners. So far, all of them are in the United States, although the company is negotiating its way back into Europe, which it left after having problems with music licenses there.
About 2 million people listen to the service on a given day, typically while sitting in front of their computers at work or, increasingly, while clutching their iPhones on the commute home, making for an average session of 6 hours. No wonder Pandora streams more data than any other site except YouTube.
Here’s how it works. The listener creates a virtual ”channel” by selecting a song, artist, or composer. If a song is chosen, the site compares it to its database of 600 000 songs, each rated by one of its musical experts. The site then selects another song it deems to be a close relative and keeps on playing such relatives. (Pandora can’t give you your first choice because its licensing contracts ban it from playing songs to order.)
When I selected ”A Hard Day’s Night,” by the Beatles, the first song I heard was ”She Loves You,” by the same band. I listened for a long time before getting my first choice.
You rate a song by clicking on either a thumbs-up or a thumbs-down icon, and the algorithm adjusts its weighting of the musical checklist it uses to select subsequent songs. What’s more, a thumbs-down will keep the channel from ever playing the same song again. You have to be careful, because the more thumbs-down you give, the narrower the channel becomes, and in the extreme case you may ”thumb yourself into a corner,” says Tim Westergren, founder of Pandora.
Even then, he notes, you’d only hobble that one channel. Nothing you do on one channel affects the others, and you may create as many channels as you want.
Westergren says he got the idea for Pandora when he was a young musician working on scores with moviemakers who had very different likes and dislikes. He wanted to find a way to encode those differences in a database he dubbed the Music Genome, paying musicians to do the enormous amount of work.
It may seem strange to use so much manpower as a supplement to computer power, but it makes sense when humans alone can handle the job--a peculiar field sometimes called artificial artificial intelligence. One example of AAI is setting puzzles, or ”captchas,” for visitors to a Web site to solve, both to prove that they’re human beings and not bots and to perform some useful chore, such as deciphering the blurred letters from a scan of an old book. Other AAI programs lure people to do such work by providing entertainment or, as Amazon’s Mechanical Turk does, money.
Westergren got seed money for the Music Genome in 2000, at the very end of the dot‑com bubble. When the bubble burst, he and his colleagues labored almost without income for five years before another injection of capital came through. Even now, Westergren says, Pandora is focused solely on growth and so does not turn a profit. It gets most of its revenue from the banner advertisements its site displays to listeners every time they click on something in the site, something they must do from time to time to prevent Pandora from going silent. It also gets a small royalty whenever a listener buys a song by clicking through to a vendor, such as iTunes or Amazon.
Pandora offers a running commentary on its songs and artists.
One advantage of using experts is that they can categorize songs that are new, by bands that are unknown. They can also provide a way to get at music that fell out of fashion before Internet rating became possible. Such too-new and too-old songs constitute a big part of the ”long tail”--the huge inventory of items that each sell in very small numbers yet collectively amount to a big part of the online marketplace. Mining that tail is one of the main jobs of any recommender system.
”In book publishing, genres are the equivalent of what we’re doing. A brand-new author can say, ’Mine’s a historical mystery novel,’ and thus put data into the product without having any customer reviews,” Westergren says. ”But our theory is that it’s not good data, not granular enough and not objective.”
Tom Conrad, the chief technical officer of Pandora, says that ”musical genomes” sometimes turn up connections you’d probably never get with other methods. He cites the ’80s pop star Cindy Lauper, who recently recorded a new record that didn’t sell in great numbers. ”We analyzed it for its genome and found that the record sounds an awful lot like Norah Jones. So we are able to play Lauper’s songs when you start a Norah Jones song. There’s a Metallica ballad that’s musically a nice fit for Indigo Girls. So start an Indigo Girls station and you might get this ballad.”
Conrad says that Pandora isn’t so proud of its expert-rated system that it can’t learn from the collaborative-filtering techniques pioneered at Amazon, Apple, Netflix, and other firms. He contends that the two approaches are complementary.
”We have benefited by peering inside the approaches tried by some of the thinking that went into the Netflix Prize competition, and we’ve incorporated some of the ideas into our own system,” Conrad says. ”I’m friendly with the Netflix personalization team; we’ve talked over the past two years or so. We wanted to have more qualitative information; they wanted more quantitative. Now we both use both. Netflix has human editors who try to capture technical aspects of the movies.”
When the two approaches meet, experts will use computers as much as computers use experts. We will have achieved the perfect chimera: a man-machine mind meld.