is the Pandora algorthm broken

Fast learner's on the track of what I'm talking about. I don't know how many of you work with relational database querying and set theory, but that's the stuff that makes pandora work.

Each song is flagged with attributes. Musical category, vocal quality, topic of lyrics.

So at some point, Band A (or songs from Band A) get flagged as:
Metal, High Pitched, D&D

Band B has songs that are flagged as Metal, High Pitched, Sailing.

Band C has songs that are Rap, Nasally, D&D

At some point, while listening to the Band A channel, it decides to only use 2/3 of the attributes and looks up bands that are also Metal and High Pitched (leaving off D&D).

thus, it exposes me to Band B and I'm happy. I'm also happy to have avoided Band C, which only has one attribute in common with Band A.


What I expected, however, is that when I make a channel for Band B (that I learned about from Pandora), the same algorthm should kick in, and eventually Band A would show up.

For purposes of explanation, whether Pandora tags by band or song is irrelevant. All songs have a band, so if it's by song, I'll still see that band show up when that song shows up as a match for attributes.

Another thing that is likely happening is that Pandora is scoring songs by how many attribute matches there are. It might be a numeric threshold (must have 4 or more attributes in common). it might be a percentage ratio (must have 75% match on attributes).

If it is percentage based, and songs have varying numbers of attributes tagged to them (rather than static categories), then it is possible that Band A has a list of 4 attributes, making it easy for other bands to only need 3 attributes. Whereas, if Band B has 10 attributes, then another band needs to match 7 of those, which Band A can never do with its meager 4 attributes.
 

log in or register to remove this ad

Experimentation in the past shows very clearly that if you start a station with one or more performers/bands it attempts to play music that sounds like the majority of that performer's work, while if you start a station with just one or two songs by that artist, the resulting music will match much more closely to those songs than the band's other work.
 

It seems pretty wonky to me some days. Then there are the cases where there's a performer that I have showing up on a channel a lot, and the info on the performer lists "similar to"s ... are a list of performers I've yet to hear a song from.
 



Remove ads

Top