Earlier quoted context omitted.
Let me try to explain it again. Suppose an integrity system has a true positive rate of 99.99%. That would be good enough to deploy right? Except that when applied to millions of accounts, 0.01% is still a massive number of people. This is even worse when those people are unusual in some way. For example they might open conversations with hundreds of strangers a day for good reasons. But their behaviour is so similar…
>"After all, there's no reason for accounts in good standing to lie, right?" No reason is different from no good reason. Also, 99.99% only seems like a high number when we think anecdotally, not statistically. For anything at Facebook's scale, the number of nines should be increased! Because .01% of the actions of three billion people on a single day gives you a city roughly the size of Tampa, Florida (~300,000 peopl…
And I worked for 4 years on one such small part (internally called UFAC), trying to help potential false positives of such a system.
As for classifier with a true positive rate of 99.99999%, I don't know much but I don't think it's possible. But if there's someone out there who might know, then they should say so.