Next Big Idea
Posted by David Harris Tue, 07 Nov 2006 05:36:06 GMT
Ok, I’ve always had this theory about music, particularly the structure behind music as a whole. Obviously, most music genres are simply a fork of an older genre. For example, we know country music is the uncle of rap music. Blues led to country and R&B, R&B leading to rap.
I’ve always thought this could be easily represented by using Graph Theory. Each genre gets its own vertex, with each edge representing how close or distant each genre is to another.
My next idea is to actually plot the whole history of music using graph theory.
My problem is I don’t know a heck of a lot about graph theory outside the 4-color theory, traveling salesman, and the seven bridges problem. However, I think I can at least get something to capture data for the graph.
Obviously, music is subjective to the listener. So I think I will devise a certain way to gather the data via social networking. Obviously, that’s the best way to get data these days.
So I’m proposing a game, similar to the old Hot or Not (or maybe that’s still around). A user gets two songs at random, and must rate 1-10 how much they correlate with each other. I think from that I should easily be able to calculate edge distance, or how well genres in general correlate to each other.
If such a graph were available, and then colored by historical era, it would be easy to spot musical trends, almost to a point where one could predict which genres will be reinvented a decade from now, and which will be abandoned temporarily.
My only problem is finding a source for such songs, and having them be accurately mapped to a correct genre.
The other problem is that the set of genres allowed in current ID3 tags (and picked up by iTunes, CDDB, etc.) is very limited for this purpose. And I want sub-sub genres involved in this, like acid jazz or reggaeton or otherwise. I’m thinking I can devise a way to collect that data from the social network too.
The benefits of this are awesome, even though it’s going to take a bit of work. Imagine an algorithm that can accurately predict what a user will like given a single correct statement up front, with given probabilities. For example, the user can say “Coldplay is my favorite band and I enjoy Baroque music”, and an algorithm should easily traverse the graph to make the claim “there is a 72.3% chance you will also enjoy Weather Report, a Jazz Fusion artist. Click here to listen.”
Such a mathematical determination would easily trump algorithms used by Last.FM and other networks to choose favorite songs based on what everybody listens to. Those are good at what they do, but it’s not very good about predicting a new emerging artist’s fan base.
At the least, it will be cool to have a graph detailing the relationships between music genres. We already know Acid Jazz is a mixture between dance/electronica and jazz, but not to what degree they are related.
Sounds like a good CS senior project, if I can delay it that long.
