Monday, March 06, 2006

Opening Pandora's Box

In response to Alan's assertion that within a user-preference network, one could simply...
"...divide each probability by the total popularity of the band in question, the Beatles quantity would be greatly diminished relative to the Children of Bodom quantity, and the preference-based Pandora would deliver us Opeth fans the truly similar Children of Bodom and not the merely popular Beatles."
Maybe I'm not fully getting your method, Alan, but it seems that someone really into British bands from the 60s would be penalized for liking such a popular band. Dividing out the total popularity would advantage the more obscure bands of that genre, where someone could conceivably simply know that they generally like the Beatles and want to hear more songs of theirs.

Yes, a program like Pandora is conceived precisely to allow access to music not normally played on the radio or universally known (read: obscure) but to me the beauty of Pandora.com is its relative lack of regard for fads and other artificial, arbitrary, human forms of classification of art. Granted, it has to use some language of classification to run its algorithm, but in this case it limits the subjectivity by breaking things down into what presumably is a more basic and objective level of analysis, ie "features."

In this way, I get to listen to music that isn't weighed down by the baggage of broad musical classifications like "indie rock" or "emo" or "nu-metal." What do these terms mean anyway? If they have quantifiable qualities that turn me on or off to them, then that should be sorted out by my feature-based preferences that I clue Pandora in on. I like not having my musical selections dictated by well worn threads through the contemporary music scene. Each suggestion thus seems that much purer, and has the potential to break down barriers between musical castes that may have been imaginary to begin with.

1 Comments:

Blogger Alan Gordon said...

Your point about being penalized for liking popular bands is well taken. Certainly with the system I implicated, if just gross popularity (p) is used, no one will ever hear the Beatles on their pandora stations, becuase p(beatles) is orders of magnitude greater than p of other bands. To get around this, I suppose we could divide preference ratings by a term that is correlated to p but grows more slowly. Something like the log of p. I retain my idea that in theory we can divide preference probabilities by some quantity related to overall popularity to acheive a quantity that diminishes the bias popularity can have in a preference-based system. How we would implicate this popularity term (i.e. how much of an anti-popularity bias we want to create) is certainly up for debate.

In your second and third paragraphs, I'm not sure why you write about genre artificiality. I agree that genres are often problematic, and I certainly wouldn't listen to a version of pandora that reduced my music taste to a set of genres, therefore giving me a load of nu-metal, K-pop, voodoo blues, and IDM that I am not interested in. What I don't understand is how genre-ism is valid to our discussion.
I think you are implying that a preference-based method, unlike a feature-based method would reinforce "artificial, arbitrary, human forms of classification of art," thus subjecting us to "well worn threads through the contemporary music scene." I disagree because I don't think that preference-based systems inherently lend themselves to genre-ism. Genre does not play a big role in how we make preferential judgments about songs. ("This song sucks. But it sounds like Nu-metal, which I like. So I'll give it a good rating") Human musical preference transcends the concept of genre which is why the arbitrary genre-ism we both are wary of is so suspect in the first place. If anything, the feature-based method lends itself to artificial categorization more than the preference-based method. Again, I cite the example of my pandora station which plays nothing but electronica. If Pandora knows that I like "synthesized beats" and "futuristic melodies" and "repetition," what genre of music do you think it will select for me?

Perhaps you think genre-ism manifests itself not at the level of individual preference, but at the network level. Again, I concede that this is quite possible, but not more so for a preference-based than a features-based system. And we can hope to reduce the genre neighborhood effect in both systems by adding a random jump component to both.

1:13 AM EST  

Post a Comment

<< Home