Musical Myware

Felix Miller, Last.fm

O'Reilly Emerging Technology Conference

San Diego, CA Mar 7, 2006

http://conferences.oreillynet.com/cs/et2006/view/e_sess/9061

Notes by Cory Doctorow

doctorow@craphound.com

--

Myware is like spyware, but it lets you spy on yourself.

Why would you spy on yourself? Why would you share the data with
Last.fm?

Last.fm: Tell us what music you listen to, anytime, all the time,
without even realizing it

Why?

Napster made all music ever recorded available -- so how do you
know what to listen to? Mission: "Harness the knowledge of the
crowd." Someone out there knows what you should be listening to;
no need to read tedious editorial.

Audioscrobbler installs in media-players like iTunes, etc and
reports what you're listening to at any moment and updates your
user-profile. Only records stuff you listen to, but not stuff you
skip -- just the stuff you pay attention to.

We process 8,000,000 submissions (info about which track and who)
per day.

Our music catalog is self-populating, through submissions. When a
new song is submitted, it creates a new catalog entry. It's
noisy. But by looking at the number of distinct users sharing a
spelling, we've assembled 8mm clean metadata records out of 25
mm.

We have the longest long-tail.

How do we datamine the knowledge of the crowd?

Collaborative filtering creates recommendations -- user-to-user
similarity and artist-to-artist similarity.

It's easy to jump from profile pages to artist pages to users,
etc. Just traversing the site creates recommendations. We also
have popularity data about which tracks by which artists are more
popular.

User profiles have your top-artist/tracks for all time and for
last  week -- what music did I love last summer, etc?

Filter recommendations from most  popular to most obscure.

Musical taste is complicated -- there aren't "rock people" for
the most part. Tags help make sense. A band is folk and indie and
singer-songwriter, etc, and we show which tags are most popular
for each item, which encourages shared tag use.

We have streaming radio -- every user's page is a radio station,
stream a user's favorite music in shuffle.

Listeners can hit buttons to thumbs-up/thumbs-down the music.

All streams are 128kb MP3s, all software is OSS.

We don't believe you should have a record label contract or
promotional budget to find the audience for your stuff.

Why should people spy on themselves? Because they get a valuable
compelling service, and get to control their data:

	* You can take your Last.FM music away as XML.
	
	* You can edit your music-stats.

This puts the listener in charge -- a social music revolution.

eof