I think in the general case you could imagine a NxN matrix of values between 0 and 1 representing similarity between N songs. Songs on the same album would have

I think in the general case you could imagine a NxN matrix of values between 0 and 1 representing similarity between N songs. Songs on the same album would have high similarity, especially consecutive tracks on the album, and different versions of the same song (e.g. live vs. studio) would also have high similarity. Then when selecting the next song, you would weight each available song's selection probability inversely to its similarity to the last few songs.