Predicting Musical Taste
A friend writes:
MIT says it has a program that can listen to a song and predict who would like it, and whether it would be a hit.I believe it in theory, although I'd love to test it out. The idea of combining it with iTunes or Amazon for peronalized recomendations sounds really cool.
Way, way, WAY cool.
MIT's Media Lab is one of my favorite outfits in the high-tech world. They regularly come up with amazing ideas and then make them work. It's unclear whether this project by Brian Whitman and Tristan Jehan (described as "MIT PhD grads") is in any way connected to the Media Lab, but it has the feel of their kind of project.
Any way, here are some thoughts on the article and the technology it describes:
It says that the recommendation engines at Amazon and iTunes "compare similarities between songs, add in the buying history of consumers, then recommend albums." My understanding is that, at least with Amazon, the recommendations are entirely based upon congruent buying histories. This is why when they launch new product categories there are brief periods of odd recs: e.g., "People who bought this also bought Clean Underwear." Or am I wrong on this?
Later, Whitman and Jehan say "it's really bad for music because it can only recommend stuff that people have bought a lot of." This, also, is untrue, as I understand it. (I don't have a PhD from MIT, though, so...) The Amazon recommendation engine only requires that (a) the person receiving the recommendation have bought (or rated) a lot of things, and that (b) there be a large enough population of people who have bought at least two things in the system overall. What makes Amazon cool is it can spot something that almost nobody is buying and bubble it up to that subset of the population that would buy it...if only they were aware of its existance and qualities. I know I've benefitted from having something obscure to both myself and the general public recommended to me by Amazon, and I can't be alone in this.
Likewise, the comments posted at the bottom of the article that are full of grim pessimism and predictions of doom and decay miss the real potential of the technology. From the sound of it, this approach takes a particular song as an input, and then makes predictions about who will like it and what they will say about it.
Yes, one of my first thoughts was, "What if you can reverse the process and have the system spit out the 'perfect' pop song?" Pop music sounds to me mostly like this is already what happens, but nonetheless I would be dismayed to hear that this technology was being used that way. However, it sounds like there has to already be public-access references to a particular song for the system to form an opinion of it at all; the web (and blogosphere in particular?) is the focus group, so if the web doesn't know about the hypothetical song, it can't say anything about it.
That aside, the real potential of this technology is for doing exactly what its creators say is their goal: "to expose the world to a wider variety of music." I have pretty eclectic tastes, when it comes to music. From a commercial standpoint, someone who just pitches down the middle is going to make a lot of money, but not any of my money. Many recent developments in technology (the one in this article is only one example) are making possible a level of market segmentation that until very recently was not even imaginable! A smart business person will use this technology to develop "micro labels" targeting market segments so small they were previously undectable.
By identifying them and understanding their tastes, highly specialized musical (and other) offerings become economically doable where before they were not; mass marketing was historically driven by the economic fact that mass media was the most effective means of communicating with markets. Mass media, however, is a hugely wasteful way of talking to small, specialized segments. The internet, the blogosphere, podcasting, massive personalized online marketplaces, and this new approach identifying musical preferences under development at the Media Lab all open up new and affordable ways of identifying and addressing the micro-segments of the market that are each small in themselves but may turn out to represent between forty and eighty percent of the entire aggregate potential market.
Far from forcing music production into an industrialized, Ford assembly-line model as one commentor laments, this and related technologies have the potential to de-McDonaldize what, frankly, has already become an over-intellectualized pursuit: matching good artists with the people willing to support their art.