July 20th, 2009

little blue dog

pandora math

Interestingly, by rating so many songs played on my default Pandora stream (which is creatively titled "Radio Radio") that I think I've crippled the probability the algorithm selecting material I haven't heard yet. In passive listening mode (i.e., providing no feedback to the program), the frequency of new selections is maybe about 1 in 1000.

About the only way I can alter what I hear on Radio^2 over an 8-hour period is to actively plant new seeds, which I typically find by starting new streams, but this takes some time to do, and having to thumb-down a lot of stuff I've already indicated that I don't want on Radio^2.

Because Pandora's rating system is boolean, another option is to rate previously thumb-upped selections thumb-down, but this means that I won't hear those selections again unless I remember to change the rating back, or use the selection as a seed; moreover, because each thumb-upped selection functions as a mini-seed, this method would unfocus the algorithm's prediction of what to recommend for me.

Another option is by exploiting what I call the snooze function, which is to direct the player to skip the rest of the selection being played, and not to select it again for 30 days. My idea is that if enough selections are temporarily removed from rotation, the algorithm is forced to come up with new stuff. Of course, one can only skip 6 selections per hour ... and snoozes count as skips ... and apparently, if you snooze a selection after reaching your skip quota for that hour, Pandora ignores the snooze directive altogether.

Which appears to mean that one can effectively stop the algorithm from suggesting new material by rating enough selections ... which more or less defeats Pandora's entire raison d'etre, it seems.

If I listen to Radio^2 8 hours a day, I can snooze a maximum of 48 selections daily. 5 days per week, and 4 weeks per month at this rate means I can snooze 960 selections before the ones I snoozed on Day 1 are allowed back into rotation. Depending on how lazy the algorithm is, unless I've thumb-upped less than about 1000 selections, snoozing at this rate (which I think is probably pretty high, on average) might never increase the likelihood I'll hear new material.

I'm curious whether this result was foreseeable, and if so, intentional ... and if so, what the point would be.