Friday, June 10, 2011

Netflix overfitting

How can you tell that I've been watching movies on R.'s Netflix account? One glance at this screen should explain it:
That's right: the top two recommended categories are "British period pieces featuring a strong female lead" and "critically-acclaimed mind-bending movies." This screen made me laugh for several minutes. Netflix is parodying my own preferences back to me with a straight face!

My comments are thus: (1) machine learning should be applied to more areas of my life. It is obviously useful, hilarious, and interesting. Also, (2) I am not terribly worried about the filter bubble so bemoaned on BoingBoing. Partially because I am aware of the automated personalization that is happening around me, I'm not worried about being trapped in a bubble with only my own viewpoints mirrored back at me. I constantly tweak my catalogued "browsing behavior" to see what sort of changes it induces in automated systems. For example, I rated a lesbian romantic comedy/thriller (season 5) as five stars on Netflix just to see how that would affect the action/adventure/scifi/British-period-...-strong-female-lead balance of recommendations. (This was also a joke on R., who didn't know I had done this and was quite puzzled by the temporary diversion of his Netflix recommendations.)

I think Netflix overfitting is an interesting case of the filter bubble. As long as Netflix's genre suggestions are interesting, I always find something I'd like to watch before I get to the point of browsing all videos, or looking one up by search. So most of the time, Netflix's suggestions are good and I watch them and like them and so Netflix is working me into an overfitted profile. Hence "British period pieces featuring a strong female lead."


This post's theme word is prolepsis, "anticipating and answering objections in advance" or, apparently, "the literary device of referring to a thing by its future state." Certain applications of predictive machine learning seem to evince prolepsis.

No comments: