2 ms·
I totally wouldn't put it past them, with the proviso that the proper claim is that they are able to predict when someone goes from "in a relationship" on FB to
by patio11 16y ago
I totally wouldn't put it past them, with the proviso that the proper claim is that they are able to predict when someone goes from "in a relationship" on FB to "single" or "in a relationship [with someone else" on FB, and this isn't perfect but it is much better than you'd expect from random chance.
For example, I think that Facebook could identity a pool of 100,000 users tomorrow and predict a breakup within the next two weeks, and be right for over 20% of them. That is statistically improbable unless you expect relationships to average about 10 weeks. (Yeah yeah, handwavy math there. Wave wave.)
How you'd do this? The same way you do any AI problem: cheat like a mofo until it looks like a problem we've already solved. For example, come up with a feature space of, say, 100 things which are easily algorithmically checkable: messaged boyfriend in past day, messaged boyfriend more than 10 times in last day, messaged boyfriend in last week, messaged boy other than boyfriend in last day, ..., changed sexual orientation listed in FB, changed residence listed in FB, ... went off to college, whatever. Assign each factor a random number from 1 to 100.
Now, construct a Hadoop cluster, grab your favorite million Facebook accounts, and use any AI technique you like (simulated annealing, GA, whatever) to play with those scores until you find a decent fit where the dot product of that vector and a vector of binary tests against a user delivers high scores for people who break up and low scores for people who don't.
Then, grab a hundred million Facebookers, take the dot product for each of them against your best performing vector, and print out the 100k with the highest scores.
This will work. Indeed, it almost can't not work, if any of your constituents of the feature vector captured any useful information. (P.S. The experimental design here is ridiculously biased in favor of success. Real life sometimes is -- it is certainly true for FICO, for example, since success is "beats a human underwriter" and human underwriters suck, what with their sucky salaries and sucky accuracy and sucky inability to process a hundred thousand applications an hour.)