The results from my second submission finally came back this morning. The server’s estimated time kept bouncing around from about 10 hours at the start to “indetermined.” I’m not quite sure how that can happen when the service is done on first-come first-served basis, but a number of supposedly determistic applications do exhibit stochastic behavior :-). At no point, however, did the server tell me that it would take a week before I’d get my recommendations back.

In any event, the results are interesting in a number of ways. Matching the eclectic nature of my tastes, the suggestions range from Billy Holiday and Frank Sinatra to G. Love and Special Sauce. It even managed to find rare artists like Django Reinhardt. Interesting, also, is that the results included Jane Monheit, but not Norah Jones or Diana Krall. Of the three, the latter would be my favorite choice due in no small part to her ability as a pianist as well as a vocalist.

Also interesting are the three songs by Chris Botti. I’m not sure where that comes from, because I’m one of those people who believes that “smoot jazz” is an oxymoron. The only music in my collection that might be considered “smoot jazz” would be one album by Lee Ritenour, and that’s Alive in LA—an album distinctly lacking in songs that one would consider “smooth jazz.”

In the comments, Steven asks how long it took me to rate my collection, and how I dealt with the “I like song X when I’m in mood Y” problem. The answer to the first is question is that it took about 45 minutes, but, then, I’m quite familiar with everything that’s in my collection. I didn’t use the method that’s described on the UICU web site. The answer to the second, is that I didn’t bother. The algorithm rather folds that into the general method by picking songs of a similar nature. As I noted above, the returned list is as eclectic as my tastes are, so the list of recommendations will likely match the various moods I have when listening to music.

As I see it, there are three problems with this method. The first is rather simple: my collection doesn’t include songs I don’t like. As such, there’s nothing that I woudl rate lower than 3 stars, and a large majority of my songs are rated either 4 or 5 stars (my “Top Rated” song list has 622 of the 880 songs in my library). This probably accounts for the Chris Botti recommendations even though there’s a very small chance that I’ll like any of Chris Botti’s music.

The second problem is more difficult to deal with. There is a common thread that runs through all of the music I like, but it’s really difficult to quantify that thread. I like music that’s multi-dimensional. For me to like it, the music needs to blend a couple of different styles in a way that brings some combination of harmonic and rhythmic complexity to the experience. Also, if the song lacks dynamics from loud to soft and back, then the chances I’ll like it are greatly reduced. I like the blues, but not necessarily straigh-ahead blues. There are some blues guitarists who are particularly adept at wringing the emotion out of a song, but, to be perfectly honest, I’d much rather listen to the Allman Brothers’ Desdemona than Eric Clapton’s version of Walkin’ Blues; or Stevie Ray Vaugn’s Riviera Paradise over his rendition of Change It. The basic emotions are the same, but the ways in which they’re evoked are vastly different.

The basic problem with any collaborative filtering algorithm is that it avoids trying to resolve questions like attempting to quantify the dimensionality of a piece of music by folding the whole issue into comparative user preferences, which leads me to the third problem. The whole point of doing this is to provide recommendations. Well, a recommendation is useless if it suggests a piece of music I’ve already heard and already decided not to include in my collection. One way around this is to limit recommendations to new music. Regardless of how we might define “new,” this limitation will significantly reduce the effectiveness of the algorithm. You run into a catch-22. For the recommendations to match any individual’s preferences, the database needs to have a large number of ratings in it. However, the more ratings a song has, the less likely that any given individual will not have heard the song and/or the artist before.

Because of this basic problem, of the 50 songs that the program recommended for me, I might actually buy one or two. None of them are songs that I’ve since listened to and decided I have to go out and buy. Based on this, we’d have to consider the exercie a failure. On the other hand, no experiment is a failure, because we at least learn something about the limitations of any given algorithm or approach. So, I still encourage people to submit their libraries. If nothing else, it will help us to figure out more effective ways to give people tools that will help them to expand their musical/artistic horizons.