20
January
2007

Failed recommendations 11:36 on Saturday

I’ve tried Amazon.com, Last.fm, iLike and most recently Goombah. All are hailed in-variously as having the best music recommendation systems, whether based on collaborative filtering, some high-tech distributed computing, or both.

Rugged Lacie hard drive

The orange rubbery thing above is my record shelf. It contains my iTunes library which has over 18.000 songs. I’m not sure if this is a problem of sheer size of collection, or just a problem of variety, but all of the aforementioned recommendation systems are failing.

My musical taste is varied, but it’s not all encompassing. The systems fail to see that I like hip hop, but I don’t prefer recent female artists who concentrate on repetitive non-melodies and booty shaking.

The recommender algorithms can’t figure out that I like old easy listening music or cocktail music, Astrud Gilberto, old bossanova, Perez Prado, James Bond themes from the sixties, but I hate new easy listening music, aka elevator music.

I might have all the albums from Prince, Michael Jackson, Janet Jackson, and Madonna. But that’s like 1% of my collection, and the systems can’t seem to figure out that owning those albums doesn’t make me an “average pop listener” who might enjoy every new average, well, semi-failed pop album.

There’s a lot of music classified as “electronic” in my library. Yet, the systems are unable to provide me with recommendations for new minimal, dubby techno, zero-beat ambient, breaks, or repetitive downtempo tracks with that certain Groove Armada vibe. All I get is some dumb-ass trance or jazzy house that sounds like failed wannabe jazz musicians playing badly, while drunk. Actually I believe the only reason for the existence of jazzy house is to get the listeners drunk.

And possibly for my interest in female folksy singers like Emiliana Torrini or Ebba Forsberg, combined with the few rock albums in my collecton, I’m now getting recommended a lot of young male “alternative rock” bands, which is a genre I find absolutely boring.

The recommendation systems fail to see there are clusters within my tastes, and fall to the trap of thinking my taste is a sum and average of its parts.

5 Responses to “Failed recommendations”

    Links from my other posts:

  1. /personal » Blog Archive » Social networking and shopping for music
  2. Comments:

  3. pil Says:

    Last.fm is a great way to find music similar to the music you like, and I know I have discovered many bands from this excellent site. It’s not the only source of information I have, but it’s another source of information.

    These recommendation systems need more data to be useful to users! I think the biggest problem in Last.fm is that you can’t tell the system what kind of music you really like. The plugin should be more integrated with the music player of your choice to get the best possible results. For example, the current 5-star rating system in iTunes does not serve any purpose other than making your own playlists, this could be a way to help the system to make better recommendations. But then again, not everybody uses iTunes…

  4. Niko Says:

    Actually I find that when I limit the data I give to last.fm, it works better. If I simply ask for similar music to artist X, I can quickly find lots of interesting new music. After my first few purchases at Amazon.com I got excellent recommendations. But when the volume and diversity of data increases, the recommendations get worse.

  5. karl long Says:

    have you checked out http://www.pandora.com yet? Does a pretty amazing job IMHO that’s more based upon the “dna” of the music as opposed to other peoples behavior. If also has some wonderful little affordences that enable you to skip songs that your tired of without saying you don’t like something.

  6. Niko Says:

    I have tried it. Actually I prefer last.fm’s “similar artist radio” over Pandora. The main difference is though that last.fm radio and Pandora play music similar to one selected artist, while iLike and Amazon try to recommend new music based on your whole music library and how the library compares to other users’ libraries.