Trading Songs

I’ve been considering a recommender system that attempts to identify individuals that are exemplars of a particular set of musical tastes. The data on clustering suggests that there are definite groupings of preference (which I’m betting map to genres). So, I’m hunting for the prototypical country listener or rap or whatever.

That idea has been done before to some extent. I know I’ve seen something on the idea of creating a sort of composite profile to represent a cluster, but I don’t know if I’ve read anything about attempting to identify an actual person who fits a cluster and then paying particular attention to that individual’s preferences.

In any case, I identify a set of exemplars. How many is a function of the data and how tight the clusters. With these people I create a system where these exemplars get positive social or emotional or financial feedback introducing new music to the system. Because their preference patterns are prototypical for the group, my reasoning is that the songs they pick will hit on whatever the key characteristics are and their recommendations will be “better” (as a function amount that a random user from the group will like the song).

Note that a convenient function of this system is that we aren’t trying to model what makes a song good even though our ultimate goal is to pick “good” songs to recommend. We are simply trying to model what makes a set of songs similar in human perception. Even if what actually makes a song good is not captured within the system, if that characteristic correlates to things that we are measuring then we can still get a correct grouping.

It’s a semantic point as much as anything, but it shifts the focus somewhat. It lends credence to the direction that Project Aura is headed because how much someone likes a song isn’t simply a function of the timbral or melodic characteristics. I know several feminists that don’t like rap. Honestly they have some pretty good arguments about misogyny and the objectification of women within that culture. A salient grouping characteristic though between a set of people and whether they like rap of not is going to be if they are feminists. That’s not the only axis, but it’s an example of a grouping that has nothing to do with the music itself that is useful if I am going to recommend a song to someone.

For each individual, a whole multitude of factors is important. Most of them are too specific to be accessible to a computer program. For example, there is a high correlation between liking “Let’s Get Retarded” by the Black Eyed Peas and having been in the Peace Corps in West Africa any time from 2005-2007, no system is going to have that specificity of data though. Many associations though are larger social constructions embodied in accessible data out there on the internet if a system can figure out how to represent and leverage them.

But to get back to exemplars and a system that uses them to make recommendations. The identification of the exemplars and their contributions don’t even have to be separate systems. I think that the financial markets have something to teach us about systems where the discovery of novel recommendations (undervalued companies) are introduced into an open market and the introducing individual gets a portion of a limited resource (the money in the world) as a reward for guessing well. With music, the commodity is new songs and the limited resource is the amount of attention people have to pay to music. At any given moment there’s a set number of listeners, and each one has a limit on how much attention they can pay, so there’s an upper bound on the whole thing.

An interesting question for further contemplation is whether to use a fixed quantity for attention (each user has one attention vote they are casting at any given moment) or if there should be a range. A range is more realistic and it might help if you wanted to tune the system to try and encourage some level of participation. Both buying music and spending time interacting with the system (behaviors I, as a marketer and developer want) is something a casual user (low attention score) is less likely to do than an active listener.

The nice thing about using a floor trading type model is the processes of identifying exemplars, rewarding them and using their expertise to drive the system can all happen simultaneously. I don’t take a bunch of songs, collect data, do an analysis, find exemplars and get them to do recommendations. I simply make a system where the strength of a user’s influence is a function of their centrality to a prototype for a cluster.

Moreso, I let them know this so they get the social feedback of knowing that their opinion is important. This will be one type of reward to encourage their active participation in the system beyond simply passively listening. People like knowing that they’re heard — I think it helps lend credence to their models of the world. It is why I wrote the piece on objectifying cognition even though I haven’t had time to tie it in yet.

And, I still don’t have time. Gotta go to class soon. There’s no telling what I could get done if I didn’t have all these irritating responsibilities. ☺

In closing though, I wanted to mention an idea sparked by Killus’ blog. If the system works as intended and exemplars are converted from passive listeners into active participants of the community, this shift could well represent a change in their values structure. As a person gets more into music they eventually start becoming music geeks, and music geeks will represent a different cluster (or perhaps combinations of clusters?). Specifically though, their preferences cease to be an exemplar for the cluster that they might have started out in.

So, the system needs to not assume that recommenders will only get better. A couple solutions come to mind. You could match them against all clusters and after some threshold is passed they start getting feedback they they’re influencing other spaces. This will likely cause them to explore those spaces and reduce their predictability, but it will draw them more toward music geekdom.

Another option is to simply add a “recommend to group” option where I can explicitly say, “I think this song would play well on this channel.” This lets a self-aware music geek attempt to maintain their finger on the pulse of a community while expanding their horizons.