Insomniac entertainment

Every now and then I go through a phase of insomnia. I haven’t been able to determine the cause. It’s not coffee — I’m actually doing a pretty good job at kicking that addiction. I don’t know. I could sure use a tranquilizer right now though.

Insomnia sure can be boring, so I’ve figured out a few things to entertain myself with. I’ve been playing around a lot lately with Pandora, an application of The Music Genome Project that I read about in the New York Times last weekend. The Music Genome Project develops a metadata vocabulary for classifying and relating music that the “music analysts” at Pandora have applied to their analysis of an insane number of tracks and artists. The result is that you can type in an artist and Pandora will pull up the defining characteristics of that artist while relating it to artists of similar characteristics. The relationships are relatively transparent too. For example, I clicked on “Why is this song playing?” when Massive Attack’s “Special Cases” was playing and Pandora related it back to my Portishead search with the characteristics, “electronica roots, downtempo influences, and tripped-out production.”

Developing relationships between artifacts using a pre-defined metadata vocabulary isn’t really a new thing (online retailers have been doing this for years), but it’s cool to see this applied to “subjective intangibles” like music artists and tracks. The difference is that developing relationships between subjective intangibles in the past has been done via consumer behaviour data mining (ie. 25% of consumers that purchased A also purchased B, so one might infer that A is related to B. Moreover, those consumers that rated A highly also rated B highly). The Music Genome Project is instead building a music classification database from the ground-up, without the help of consumer behaviour data.

There are merits to both approaches: mining consumer behaviour data is probably a lot cheaper than analyzing immense numbers of artists and tracks for their defining characteristics, for example. But one might criticize relationships derived using this method as being predominantly related to economic behaviour rather than defining musical characteristics. Consumer behaviour and preferences for certain musical characteristics must be related to eachother in and of themselves, but still, the data mining approach seems a bit flawed in that it doesn’t account for these relationships based on musical characteristics.

Maybe the silver bullet would be to combine data from the two different approaches? One could overlay relationship data derived from consumer behaviour data mining with the relationship data from Pandora, reconcile it, and come up with “meta-relationships” based on the two sources of data. It’s an idea. An insomniac idea.