inTune With Me

I've recently begun a literature search in the space of "recommendation engines." These are systems that leverage information about you (and perhaps others like you) to make a suggestion about things you might like; think amazon.com's recommender.

Quite honestly, I can't say I have ever been truly impressed with the current crop of recommenders. Though, admittedly, I probably haven't used that many of them -- perhaps out of lack of opportunity. However, in the fall of last year (at the beginning of my first term at UW) I subscribed to the Yahoo LAUNCHcast music service (~$3 a month). Note that LAUNCHcast also has a free service with a limited number of radio stations. Song skipping is also crippled (you can, i think, skip 5 songs in a day rather than an unlimited amount). You also then have to listen to a few short commercials every hour.

Launchcast's tagline is "the music that listens to you." Using a method called collaborative filtering, Launchcast introduces new music into your customized radio station that the system thinks you might like. How does it work? The success of the recommender system rests largely on the willingness of the user to actively rate songs. Song, album, artist, and genre ratings are combined with the existing launchcast userbase to find "similar listeners." These listeners have roughly the same ratings as you. Launchcast then begins to play songs that they have rated highly but that you've not rated (or perhaps heard).

Users will be more likely to participate if they feel like the value returned is high. However, the return value is rather low until you (and others, as the system uses collaborative filtering) begin rating songs. This is called the bootstrapping problem -- an issue that exists in many recommendation systems.

I don't mind rating songs. I rate songs in my iTunes library all the time so I have a refined approach to rating songs on launchcast. In addition, music is my passion. It is one of the few things in life that I freely admit to loving without flinching. So, the work that I do to rate songs is worth it to me because I will (hopefully) learn of new music that I like in return.

And, in this case, it's actually worked. I would say that a fair amount (perhaps as much as 60%) of the new music I hear on launchcast, I like. A subset of that (~15%), I really like. In fact, if you look at the music that I have purchased in iTunes in the past six months, a majority of it was introduced to me by launchcast. That fact alone has enormous implications for e-commerce. Suddenly, technology has taken the role of my "trusted musicphile" friend who sends me mix tapes in the mail; though I'm not sure it is yet an adequate replacement.

Of course, the iTunes music store itself has yet to really take advantage of user's metadata. You could imagine iTunes scanning your music library to derive your likes and dislikes (perhaps as a combination of playcount, ratings, and simply number of songs by an artist) and suggest new songs that you might want to buy. They do, however, make recommendations to you based on what is currently in your cart. This obviously has shortcomings.

Lately I've been imagining an iTunes based online dating service where you pair two clients together based on their music tastes. Or, perhaps something more benign, like a friendster that connects people together into the same social network based on similarity clustering of their music libraries.

Still working on what I've dubbed "inTune." A friend yesterday told me the name sucks. Perhaps it does. Not that it matters at this point.