Tuesday, February 14, 2006

iTunes music recommendations

Alyce Lomax at the Motley Fool notes that Apple's iTunes Music Store is now making music recommendations based on past purchases.
The iTunes Music Store has a beta feature called Just For You, which suggested albums I might be interested in based on past iTunes purchases.

Was it accurate? You bet. I bought a few albums, EPs, and singles.
Alyce goes on to say the feature is comparable to Amazon.com's recommendations and that "accurate recommendations add much more than convenience to Internet-based shopping."

When I fired up iTunes and took a look, my recommendations seemed reasonable. In my case, the recommendations overemphasized some older purchases that no longer reflect my tastes in music, but that's probably a minor issue that they'll iron out soon enough.

I'm pleased that Apple is doing this. More than once, I had enough trouble finding things to buy on iTunes that I've hopped over to Amazon to look at their similarities and recommendations, then returned to iTunes to buy. Of course, most people wouldn't bother doing that.

It's about time Apple added more personalization and discovery features to the iTunes Music Store. Let's hope there's more to come.

[Found on Findory]

Update: The WSJ reports that Amazon.com soon will offer iTunes-like digital music downloads. Maybe I won't have to wait for Apple. [via TechDirt]

Update: Kate Moser at the CSM just wrote an interesting article on music recommendations. It includes a prediction that "taste-sharing applications" will drive 25% of online music sales by 2010.


Anonymous said...

I heard a rumor that these recommendations are powered by Choicestream. Have you played around with Yahoo! Music recommendations in the Yahoo! Music Engine? These are based on ratings as opposed to purchases. I'd be interested in hearing your comparisons.

Greg Linden said...

Hi, Todd! Great to hear from you!

Interesting, I wouldn't have guessed that Apple turned to Choicestream for this.

If so, that means that the recommendations are content-based (using metadata about each song or album) rather than based on user behavior (using what other people tend to buy). Choicestream uses a content-based approach they call "Attributized Bayesian Choice Modeling".

I'd love to compare to Yahoo Music and Yahoo LaunchCast but, unfortunately, that's really hard to do unless iTunes adds ratings so I can experiment with different profiles.

Thanks again, Todd, for letting me know about that.

Anonymous said...

At the risk of being somewhat of a pedant, I just have to say again that "content-based" music retrieval is not the same thing as "metadata-based" music retrieval.

"Metadata" retrieval would be like searching tags, or annotations, of the music. "Content-based" retrieval means matching based on the actual musical signal/audio itself.

I looked at the choicestream website, and their PR brochure, and they said the recommendations they provide are based on a "deep understanding of music". However, they offer no insight into what that "deep understanding" entails. Can they really, for example, automatically distinguish between a clave son and a cua polyrhythm in a salsa song? Especially when no one has explicitly "metadatacized" this information by annotating or tagging the song with this rhythmic label? Can they go into the music itself and "hear" the differences?

Or is it just "this song is pop/rock, this song is R&B, etc., because someone at CDDB labeled it this way"?

I guess my only point here is that there are actually three main sources of data for musical similarity algorithms:

(1) User behavior (playlists, buying patterns)
(2) Metadata (i.e. annotation or tagging -- attribute-labeled data)
(3) Content (similarity in the actual musical signal)

Most people are doing #1. Some people (Pandora, and maybe also Choicestream) are doing #2. Academically, people have been doing #3 for 6-7 years now, but I've seen no commercial applications of it yet.

Unless you know more details on Choicestream? Are they actually delving "deep" into the music itself, as they claim?

Greg Linden said...

Hi, Jeremy. I'm sure Choicestream was using the term content broadly, as was I, to mean text data associated with the music.

You have a good point that it is important to distinguish between analyses that look at the music stream and those that do not.