Tuesday, December 07, 2004

Music Recommendations

For about 6 months now i've been working for one of the on-line digital music services. My personal taste in music is eclectic, which indicates either curiosity and open-mindedness, or a complete lack of discrimination. One of my main motivations for taking this job (my previous career was almost exclusively in scientific domains) was to have access to the music. Since i work in the "personalization" part of the business-- meaning that we use various techniques to determine the music that a person might like based on what they're already listening to-- i figured that i'd find all sorts of obscure artists that i'd never heard before.

Well, i did discover many obscure artists that i'd never heard before; but unfortunately that doesn't mean that i enjoyed listening to them. I managed to hear two new bands that i like-- Snow Patrol and The Dresden Dolls-- prior to their debut on local radio. (The Dresden Dolls record is excellent btw, easily my favorite of the year. They write the best lyrics i've heard in ages, and the music is unique yet still very catchy). I've also been able to listen to much more of the Wilco discography, and i've listened to a few of the related artists that i probably wouldn't have purchased otherwise (Whiskeytown, Uncle Tupelo, Ryan Adams). I've sampled a few of the hard-rock bands related to my personal favorite, A Perfect Circle, but the only band that i've listened to more than i probably would have otherwise is Sevendust.

But a significant majority of the artists that i've checked out as a result of an automated recommendation are obscure for good reasons. Presumably these bands have followings, since they got record deals, and it's great to have the option of listening to them on demand rather than purchasing a CD. But you'd hope that the technology would lead you to stuff that's really good and just didn't get airplay. This, after all, is the idea behind The Long Tail.

The problem with recommendations technology in music is, i think, the flipside of what makes it desirable in the first place. If you're a fan of a particular band, then it makes sense that you'd want to hear music by similar artists. But chances are that if you're a fan of a particular type of music, you already know most of the similar artists. For example, i love Duke Ellington and our service provides 94 recommended artists related to Duke Ellington. But because i'm a fan of jazz and of swing in particular, there isn't a single one of the 94 recommended artists that i haven't heard of before; and i have to go to number 38 before i find an artist whose music is not familiar to me (Bill Evans). Of course, the results are different if you start with less popular artists. For example, i like the band Giant Sand and their associated recommendations are less familiar. The top two recommendations are Cat Power and the alternative country band 16 Horsepower. Both artists are very listenable, but neither are something i'd pay money for. In the case of less popular bands, i think the peculiarities of personal taste are too hard to account for.

Books are different, i believe, because the culture of books is less fan-based and preferences are less style-based (although, i'm sure that virtually every literary critic who knows anything probably thinks i'm an idiot for stating the above). I use the Amazon recommendations fairly often. Often this is for technical or scientific books, but even with fiction i've had good success. I've bought a couple of books based on recommendations related to my purchase of books by Michael Martone, and i've been pleased in both cases (Notable American Women: A Novel by Ben Marcus, and Hideous Beauties by Lance Olsen). There's enough similarity that you understand the recommendation, but not so much that you feel like you're spending too much time in the same territory.

My main conclusion from this experience is that the existing music marketing machinery doesn't suck as badly as i always assumed it did. Pinback's Summer in Abbadon, another favorite from this year, i first heard on the radio via the single Fortress. I got into Kanye West after reading an article in the on-line publication PopMatters. I heard the new Chevelle record on the radio before i noticed it on-line, and i first heard Coheed & Cambria on Fuse! (Sure, i'm ashamed to admit that i like Coheed&Cambria and that i watch Fuse). I remember reading some on-line article a couple of years ago that argued that the record companies are valuable for acting like a filter for musical taste rather than for production and distribution capabilities. The writer argued that as soon as other filters became available (i think he promoted Bayesian statistics), the record companies would cease to be relevant. Well, one can hope, but we're not there yet.

1 comment:

Greg Linden said...

Do you think the problems you're seeing are inherent to any music recommendation engine? Or just the particular music recommendation algorithm used by your firm?

You said you like the Amazon.com book recommendations. What do you think of Amazon's music recommendations? You'll have to have bought or rated 5-10 albums on Amazon for the recommendations to be any good, but I'd be curious to hear if you find them more useful.