That would be very interesting (to me, at least). In particular, I'd be curious to know how you test your inferences. I guess you could run it on the NetFlix/GitHub prizes ... but in general, I'm finding that getting reliable data on which to validate algos is one of the biggest challenges.
In any case, I'd be interested to hear about the math, since there are about 50 different ways of doing this stuff. Even just vague stuff like "We might use a boltzmann machine".
There's some interesting work done on optimal stimulus selection (MacKay has a paper on it, and there's one in a neuroscience setting by Paninski). The idea is that you generate data for which the response will give you maximum information. So you figure out which edges would be most valuable to your algo's predictions (using information theory) and then make recommendations based on hypotheses about those edges, which are later confirmed/denied by user behavior. This gives you an optimal learning loop.
In any case, I'd be interested to hear about the math, since there are about 50 different ways of doing this stuff. Even just vague stuff like "We might use a boltzmann machine".
There's some interesting work done on optimal stimulus selection (MacKay has a paper on it, and there's one in a neuroscience setting by Paninski). The idea is that you generate data for which the response will give you maximum information. So you figure out which edges would be most valuable to your algo's predictions (using information theory) and then make recommendations based on hypotheses about those edges, which are later confirmed/denied by user behavior. This gives you an optimal learning loop.