June 4th, 2007
NYTimes has an article on the ongoing Netflix recommendation open challenge thing. ($1mil to a team that can produce the best collaborative filtering mouse trap)
Unfortunately the project is flawed, because the basic question is flawed, fundamentally and in a very simple way. We have moods, we have shifting interests, and trying to compile all those multi-variates into a single vector of interest is impossible.
Rather then making the computers super smart, I’d rather see an interface like the Pandora channel creation where you choose 2-4 songs that suit your mood and the system finds the common elements.
Tonight at the video store I wanted something that was smart and fast enough to be engaging, without being so smart that it took work to follow. Maybe a political thriller? If I could have mixed a recommendation queue out of 3 Days of the Condor meets Wag the Dog meets something like Enemy of the State to find something in that vein that would have been better then all the weighted neural nets.
And I’ll wave the million dollars if they just build it already.
Not sure where this systemic biasis for computer as deep thinker comes from, probably dates all the way back to the Ultra project and other primordial computer science legends. But its the wrong metaphor here and now, smarter, smaller tools to extend the human reach, not replace humans.
(can you tell I didn’t find my movie?)
Photo by wili
January 21st, 2006
The post on BB today reminded me of MusicBrainz. I’m still patiently waiting for MusicBrainz to add support for arbitrary, user-contributed, emergent flat classification. (i.e. what the rest of us calling “tagging”, but that term is overloaded for audio files.)
My original desire was driven by the desire to build playlists by mood, or theme (‘electronic’, ‘sensual’, ‘dark’), but since then Pandora has launched and showed us another set of axes to explore (‘mild rhythmic syncopation’, ‘extensive vamping’).
I think the ability to hang a variety of arbitrary data off of the MusicBrainz model would kill CDDB deader then a stake through the heart, and at the same time creating a platform where its easier to collaborate in public then in silos. Tagging is just the first, easiest to under form that might take. Got to use all that lovely RDF to some purpose!
Tom’s Phonetags is useful prior art.
January 7th, 2006
I’m getting good stuff out of Pandora mixing Frontier Psychiatrist with Feel Good, Inc. What are you mixing?
Recommendation cocktails are the way to go.
Most recommendations are either one dimensional (if you liked X you’ll like Y), or, more often, assume that all our little quirks added up describe our one true nature (e.g. Amazon). In fact we’re more complex then that, described by a multitude of often unrelated vectors. Pandora lets you experiment with the dot products.
Thanks Rob for pointing out this feature, I had missed it the first time through, and had written off Pandora.