We Can Fix That with Data

I’ve been using Pandora a lot lately. Pandora is a music matching service — you enter a band or a song, it does some fancy queries, and spits out more music with similar qualities. You can “thumbs up” or “thumbs down” each song to refine the qualities you’re looking for. If you didn’t like it, it’ll go on to the next song.

They don’t hook you up with other users’ preferences — data mining is for the weak. They manually analyze every track and look up your stuff in that database.

. Together our team of fifty musician-analysts have been listening to music, one song at a time, studying and collecting literally hundreds of musical details on every song. It takes 20-30 minutes per song to capture all of the little details that give each recording its magical sound - melody, harmony, instrumentation, rhythm, vocals, lyrics … and more - close to 400 attributes!

The qualities they look for even vary by genre.

Rap, for example — there’s more detail around lyrics in the rap genome than there is in the pop genome because rap is so much more lyrically focused. We’d focus on the literary and delivery — rhyme schemes, rhythm, and wording; like how much profanity there is.

I think the last quality here is the reviewer getting tired.

Interestingly, I got a promo email yesterday that hinted they might be branching out into the brave new Web 2.0 world: “our new features help you find new music through other Pandora listeners, and let you tell the Pandora community about your favorite music.” Is manual labor getting too expensive, or are they trying to get into the friends business?

Here’s my new profile page, where you can browse my stations and see what I’ve thumbed up and thumbed down. You can search for other users’ stations as well, and when you look at an artist’s page, you can see other people who like them (along with a parenthetical note about other stuff they like). For example, a quick glance at the Zeromancer page shows that people who like Zeromancer tend to like other stuff I don’t.

I always screw up my stations — even though each one is based on a different sound, they tend to merge over time. This is because I thumbs up artists I like, even when they don’t match with my goals for the station. The thumbs up doesn’t actually mean “I like this artist,” mouseover notwithstanding; it means “I want more of this on this station.” Pandora doesn’t tell you this until you notice they’re playing KMFDM on your Kovenant station because you said you liked Zeromancer. I like all three bands, but I want to keep the electronic industrial on the KMFDM station and the industrial metal on the Kovenant station, you know? Now I either have to move songs from station to station, a pain, or I have to thumbs down bands I like, and I feel dishonest.

They could verify whether or not other people have that problem — out of users with more than one station, how similar are their choices, and do they become more similar over time? I would think that simple usability problem would be problematic for other people too. On the other hand, changing the mouseover would add more words to the interface, making it feel more complex.

So how do they plan to make money?

Approaching the question from a theoretical standpoint, Pandora’s potential commercial success is based on an economic concept called the “Long Tail.” In the digital age, where consumers can easily find the specific thing they’re looking for and producers can easily provide new content for distribution, Long Tail states that directing consumers off the beaten path is a potential key to profitability.

I’ve already spent a chunk of cash on bands I’d have never found otherwise. Some have been good (I’m really liking the Kovenant), some have been bad (I’m embarassed that I spent money on that horrifically awful Deadstar Assembly album). It’s still great to learn that the last 10 years have actually produced music that I’m interested in, and it’s interesting that they’re telling me about it with old-fashioned man-powered methodology.