The talk was good, as all of them have been, though maybe a bit less organized around a central theme. There was quite a bit of dabbling in advanced music theory (that left many attendees scratching their head) as well as some very accessible music playing, from Coltrane to Zeppelin to Pachelbel. Clearly, Glasser possesses a unique genius.*
Amidst the offshoot thoughts spurred by the talk were some audience questions about predictive analytics. Pandora is, at root, a big pop culture analytics experiment. Each track is scored on a scale for 100s of attributes. There are millions (billions?) of tracks. Those are a lot of data.
Three thoughts on these data and how they are analyzed:
1. The human element has not been automated.
Each Pandora track is coded by a human being. I found this really surprising. And sort of gratifying. I remain suspicious of (though optimistic about) highly automated analytics, especially social media monitoring and natural language processing programs. That the Pandora model, especially given its mathematical and theoretical basis, has not completely handed the data inputs over to an algorithm made me feel better about my suspicions.
2. There are critical differences between things that can’t be predicted…
Someone asked if the model could be adapted to predict future radio hits. Gasser was doubtful. It’s a great idea, and maybe it is possible, but it’s important to be realistic about the limits of your data, your algorithm, and your knowledge of the human framework that underlies all our decision-making.
3. …and those which can be predicted.
The beauty of Pandora radio, of course, is that it predicts what songs you might like and adapts its model based on your inputs. Pandora works not simply because it has a massive, static database but because that database is modified with individual inputs.
Another place where Pandora succeeds is in measuring its audience analytics. Gasser mentioned briefly at the end that Pandora has really redefined online marketing with its ability to deliver targeted advertising. I would love to hear more on this topic.
(Aside from analytics, this talk opened up a whole bunch of questions for me on aesthetics. Gasser works from the premise that we are biologically hard-wired to like music, and that there are mathematical properties that make us like some music more. Yet he is unwilling to reduce taste to an equation. I asked him a question about this, not very well, but there was one piece of his response that I particularly liked. Basically: “We’re all hard-wired to laugh, too, but each person’s laugh is different.” I thought this was an interesting take on the matter of taste. But I’d love to ask some follow-up questions on this topic as well.)