Archives for July 2006

This post will (temporarily, at least) close the loop my recent discussion of good music prediction systems.

One service that initially escaped my attention was Last FM, aka AudioScrobbler. Perhaps it went unmentioned because it’s a bit of a hodgepodge when it comes to features – it tracks what you listen to, but compiles only the vaguest (and in my experience, often incorrect) statistics about your listening habits. It features some free music, but not in a predictable enough fasion that i’d use it on a regular basis.

Since it doesn’t accumulate anything but playcounts, Last.fm can only predict based on your listening habits. For someone like me who listens to 1k+ tracks a month, the approach is fascinating but ultimately scattershot, as it isn’t weighting my likes and dislikes at all. Though it has the plus side of offering predictions based on a large network of users who you can either friend or “neighbor,” the lack of any rating scheme is a major turnoff.

That said, i return my attention to Yahoo’s LaunchCast Radio.

I have been phasing this out at work now that I have a new iPod, and it’s unaccessible at home since it doesn’t work in Firefox. However, i remain convinced that it comes the closest to being the best music service out there based on the strength of its predictive abilities. It has lead me to more than a few downloads and purchases in the last month, many of which have been surprisingly obscure.

I definitely recommending trying the service, and do so with the following recommendations:

• When you first subscribe spend a day or two listening to one of the pre-set stations that’s nearest to your tastes in order to give the service some ratings to work with. Alternately, take a sampling of your record collection and add 200-500 ratings – probably enough for the services correlative powers to kick in.
• Unless you enjoy a *wide* swath of music in one particular genre it’s in your best interest to rate genres very conservatively, especially high-level buckets like “Rock” or “R&B.” Rating “Rock” highly partially thwarts a rating of “Don’t Play” for “Classic Rock.” Furthermore, the system seems to prefer genre recommendations to song correlations, which is increasingly frustrating as you fine-tune your song ratings. Just as bad, if not worse, if you leave genres blank Yahoo assumes you like them all equally!
• One positive impact is that if you have a subgenre you’re interested in hearing more of, like “Big Band” or “Zydeco,” you can rank it heavily for a few days to get served a bigger sampling of songs so you can develop your opinion.
• Similarly, only rate an artist if you want them to impact the system’s choices. You might love Madonna or Depeche Mode, but if you aren’t interested in the terrible pop they’re correlated to you might be safer just rating songs and albums. Rating a smattering of songs by an artist has an equal (or better) effect on being served more songs by the same artist as rating the artist themselves.
• Whenever you hear a song you really like, click the song name to view its entry, which contains its similar songs. This is especially fun when listening to classic music that you don’t necessarily own, as it tends to jog your memory for other songs you’ve forgotten. (When you hear a song you really hate you should do the same thing; you might kill ten terrible-sounding birds with one well place stone. Or, you could find out a song you love is too closely correlated to the distasteful pick).
• There’s a fixed amount of time (or number of songs?) you can consume in any given month before higher features are locked out, leaving you only with your own station with a somewhat limited pool of songs. Our office seems to hit this point about 2/3 into a month. If your tastes run mainstream the limited pool is actually not so bad, but to avoid this make sure to shut (or at least pause) the player when you leave your desk.
• Though Yahoo’s awesome correlations per Artist, Album, or Song help support predictions of your taste, the system seems to be incapable of adapting to a non-standard correlative scheme on a per-user basis.

  For example, what if I rate “Don’t Play” on every song over five minutes? The system would learn to avoid long songs that were similar to each other, but voting no to Queen’s “Bohemian Rhapsody” and LedZep’s “Stairway to Heaven” wouldn’t necessarily protect me from Fiona’s “Never Is a Promise” or Tori’s “Yes, Anastasia.”

  A best-of-class predictive system would be able to determine your tastes not only based on correlated predictive data, but also based on your personal trend of ratings for certain song lengths, BPMs, producers, labels, or even mix levels and/or frequency response.

If you know of this promised land of music consumption, please point me in the right direction. Heaven forbid i learn enough programming/scripting to be dangerous, i might have a go at my own datamart, a la iTunes Registry.