The music industry forces fans into neatly-packaged "genres" -- you're a "pop fan" or a "country fan" or a follower of alternative rock. This, in turn, is how our music is packaged: Turn on the radio and you'll find a "rock station" that regurgitates the same tunes you'll find on any other "rock station" anywhere in America. And if you're a "rock fan," this station is supposed to satisfy all your musical needs, sending you running to stores to purchase, lemming-like, the albums fed to you by some program director paid off by record labels to promote their albums.

But musical tastes often run wide instead of deep -- which makes it difficult to recommend music to a stranger. Unless you are intimately familiar with a person's musical tastes and quirks, suggesting new music is often like trying to shoot at a saucer with a rifle full of buckshot; sure, you may occasionally hit, but just as often you're going to miss your target altogether.

In 1996, MIT Media Lab professor Pattie Maes decided that the best way to tackle this problem was through collaborative filtering, and helped launched Firefly's BigNote music service. The concept was neatly logical: users would rate and review music, building a grand cross-referenced database of musical tastes. The more you told the system what you liked, the more Firefly would be able to make specific music recommendations based on what other similar Firefly users liked. If you rated Alanis Morissette highly, and other Firefly users who rated Alanis highly also showed a fondness for Fiona Apple, Firefly would turn around and recommend that you check out Fiona Apple. Elementary.

The collaborative filtering concept very quickly became the buzzword du jour. Companies like LikeMinds and NetPerceptions sprang up overnight, offering collaborative filtering services to Web sites that would help you find everything from good books (such as Amazon.com's Recommendations Explorer) to movies (like MovieCritic.com) to Web sites (such as SiteSeer.com). But most of these sites faded quickly as the collaborative filtering buzz gave way to the next hyped Net hysteria. (Push, anyone?) Firefly searched for a business model and floundered; the pieces of its technology were sold to Microsoft and Launch.com in 1997.

Firefly's music recommendation engine was more of a novelty than a truly useful tool: Sure, it could tell you that if you like Alanis you'd like Fiona, but wasn't that pretty obvious already? The system would rarely, if ever, break out of the mold of mainstream bands and recommend fringe music you'd never heard of before. And if your tastes strayed across numerous niches -- say, you liked country and pop and techno, but weren't particularly devoted to any one genre -- Firefly was equally problematic; the odds of finding a community of users with identically eclectic tastes were slim.

Maes was aware of the limitations of the system. "One big problem that Firefly didn't solve is that when all the recommendations are only based on what other people enter, then it's harder for new artists to get recommended," says Maes. "Enough other people need to have taken an initiative to put in a rating for an artist before it gets recommended to people with similar tastes."

Maes now serves on the advisory board of Media Unbound. And while Media Unbound uses collaborative filtering at its base, the system also relies heavily on the opinions of critics. Not only does the software use statistical models and mathematics, but Media Unbound's staff of musical experts pick out the best music and build extensive maps with "distance metrics" that correlate how close to each other various genres, bands, sounds and songs are. These critics would, for example, recognize that R.E.M. had several different "periods" and that songs from specific eras appeal to different tastes; if you don't like "Shiny Happy People," you might still like "Exhuming McCarthy."

Before offering any advice the Media Unbound software asks questions like "If you were talking with a friend, how would you describe your music preference? Sound, Artist, Mood, Era or Genre?"; "Name your favorite five bands"; and "What type of guitar instrumentation do you like?" Next, you'll rate clips of songs from three of your favorite musical genres, and explain what, exactly, you like about the music. The resulting recommendations are surprisingly precise. After fifteen minutes, Media Unbound will create a personalized radio station for you, using some music that you've heard of and a lot that you haven't; but almost all of which you'll like. The more you listen, the more the system will tailor your personal profile.

"In general, we try shying away from drawing distinct lines," says Papish. "Saying this song is alternative, and therefore you are an indie listener so you can't listen to that song. The music universe is very fluid, so we can draw parallels and lines around things and let you move around them, but we don't limit you to a particular bucket."

I, for example, am very picky when it comes to indie rock; I don't like much of it, but do have very specific bands that I love, which cross a number of genres. The radio station that Media Unbound put together for me was perfect: It had the Pixies, Liz Phair, Massive Attack, Sleater-Kinney, Stereolab and Elliott Smith, some old Cure and Smiths classics, certain Nick Drake songs, and a bit of Belle and Sebastian -- all of which perfectly epitomized my rock tastes. But it also spit out artists like Gomez, Helium, the Lucksmiths and Neutral Milk Hotel; obscure bands I had never heard of before, but which I discovered that I liked. (Media Unbound also suggested a few songs I hated, but heck, no technology is perfect.) In just half an hour, Media Unbound managed to convince me to investigate 10 new bands that I never would have heard on the radio.

The Free Software Project
Read Andrew Leonard's book-in-progress on Linux and open source -- and post your comments.