[<< | Prev | Index | Next | >>]

Thursday, November 02, 2006

Netflix Update: Don't Try This at Home

[Followup to this.]

I finally broke down and stopped obsessively running the train/probe set in advance of training up the full set for submission since by definition it takes twice as long. I had wanted to get a feel for where it would start over-fitting, but my lil' laptop can only do so much each night churning away at the foot of my hostel bed. So this next submission (which should be up by the time this post is) will be a blind one, with no idea how well it will do. I sped the implementation up by, oh, a factor of 10 or so, and just counting training time, this submission was raised from scratch in 15 hours. Drum roll please... (watch it be worse than my last one, heh). [Edit: Heh, ok, so now I'm number 4. :) ]

For those who are scratching their heads over my algorithm, I should clarify that I did do some substantial re-derivation vs. what is in the paper I referenced. The overall approach is similar in spirit, as is the implementation (close enough that I called it a tweak), but there are some important differences in the math. So, don't expect that paper to fit the problem as-is--it doesn't.

For those who emailed me, I do intend to get back to y'all, but my online time is quite limited at the moment so it may take a while. Also, since the question has come up a lot: Yes, I would love to publish and/or open-source my algorithm, but not yet as I don't want to burn any bridges. (Incidentally, rather like the Nips Reject entry, most of this stuff would have been published a long time ago if the academic publishing industry wasn't so hoity toity. And I mean a long time ago...)

[<< | Prev | Index | Next | >>]

Simon Funk / simonfunk@gmail.com