Netflix: Actual Thoughts
My "Mechanical Turk" reference in the previous post was, as Michael Chermside says, a facetious strategy of using Amazon's cheap-human-labor service to improve movie recommendations. I've been trying to come up with a joke about this that involves the word "Bollywood" and yet which isn't potentially offensive...
In the real world, I have been impressed in the past few years by a couple machine-based learning systems: the handheld 20-question gadget driven by the algorithms at 20Q.net and Pandora.
The 20Q.net knowledge was apparently derived from a Bayesian network. (Real Bayesian networks, not the simple pivot used by Bayesian spam filters. Also, search for "median of medians")
Pandora, on the other hand, works via expertly-defined, human-assigned characteristics that have been applied to a large set of music. For instance, one of my radio stations is emphasizing "a subtle use of vocal harmony, mixed acoustic and electric instrumentation, major key tonality, and many other similarities." Now, that's not how I'd describe The Velvet Underground (the band with which I started the station) or The New York Dolls (the band currently playing), but however they did it, it's an appropriate link.
Another thought on movie selection: one of the things that really plays into movie selection is "I can see why you'd think I'd like it, but I hated it." For instance, American Beauty was obviously an attempto do worthwhile things, but I despised its weak-headedness. Jim Jarmusch seems to always fall a little short of his potential, but I'd prefer watching his worst work rather than spending two hours with Kevin Costner as a Coast Guard rescue swimmer. (For that matter, there's never been a movie -- not The Perfect Storm, not Master & Commander -- that's come close to communicating how frigging scary an open-ocean storm is. So any time I see a movie that has blowing foam and a dump tank throwing a knee-deep wash across across the deck, I subtract half-a-star right there.)
One of those no-conclusion posts...