For Pandora, they use a vector distance algorithm (with some other nudges to keep it from being repetitive, I'm sure). They break down a song on a scale of 1-5 across hundreds of "genes" and then pick songs based on how similar or "close" they are in the multidimensional space.