How Does iTunes' Genius Feature Work.

Factual Questions
1

I assume the underlying code is a secret, but I’m interested in the basic theory for how iTunes uses the cumulative ratings of millions of user’s songs in the music library to recommend appropriate songs in a similar genre.

ITunes’ approach seems fundamentally different form Netflix’s recommendation which assumes that users who have a large number of similar likes and dislikes in common will have similar taste in unrated movies. Unlike previous recommendation algorithms, the genre of the movie is unimportant; it won’t suggest the “western” Blazing Saddles just because the user liked Shane and The Searchers.

With, iTunes, however, the genre, or at least type of song (loud, rap, oldie, etc.) is more important. If I select Peter, Paul and Mary’s “Leaving on a Jet Plane,” iTunes’ Genius recommends Bob Dylan’s “Blowing in the Wind.” How does it know that the two songs are musically a good match just because many people may have rated them the same? I’ve also rated a lot of hard rock, alternative and grunge songs highly. Why doesn’t it also suggest Nirvana’s “Smells Like Teen Spirit” since I’ve given it five stars, too. I’m sure I’m not the only one with all three of those songs both A) in my music library and B) with five star ratings.

2

Actually, I’d suggest it probably uses iTunes purchase data as the underlying source, rather than ratings.

3

That’s what Amazon’s book-recommending feature does.

There’s a website called Pandora which allows you to stream songs online in the style of a particular song or band of your choice. It works using something called the “Music Genome Project”, which apparently involves in-depth analysis of each available song to find commonalities.

It then tells you what common elements it will use based on your song/band choice to select similar songs. A lot of the elements are pretty unintelligible to laymen, though; here’s an example (using Sepultura’s Refuse/Resist):

4

It seems to pick songs at random, not out of any other algorithm. There is a vague connection, I suppose (if you pick a blues song, you’ll get blues songs), but it’s not particularly sophisticated.

5

What’s odd is that it seems to aggregate all the features of a given band’s songs together, so if you pick a band like Faith No More it will get really confused- the first three songs it picked when I did that were Enter Sandman, Extreme’s More than Words, and something by Duke Ellington (which I assumed it decided was related to I’m Easy, a ballad by Keith Carradine that FNM did an ironic cover of)