If I were doing this, I'd be interested in finding ways to find out more about what people like and don't like in the generated songs and then use that to improve the algorithm.