8:45 AM
@Makyen At the risk of beating a dead horse, I want to make one more attempt to reply simply/clearly regarding your earlier statement "I'm saying that if you have total posts = #TP + #FP, then with a fixed %TP, the ratio of #FP to #TP isn't going to change, because the %TP describes that ratio".
To be clear, the %TP isn't fixed globally, it varies according to the threshold used, and that threshold impacts the fraction of the total available posts (all future posts with scores >= 1). So you can "Fix" the %TP ratio of autoflagged posts, but the number of autoflagged posts increases.
Using the simplified numbers, while 10% of 100 is 10 and 10% of 200 is 20, the 100 and 200 in that example aren't the total number of posts. They are a subset of all posts (which you note in your later reply, with caveats about future, randomness, etc. with which I agree).
So there could be, in your original example, 1000 total (future) posts that the system evaluates in X timeframe, 500 of which are tp posts and 500 fp, and we are only autoflagging 100 at the higher threshold. At a higher threshold, only 100 total posts are considered for autoflagging, with that 10% rate meaning 90 tp + 10 fp + 410 tp not autoflagged + 490 fp not autoflagged.
Without changing anything, if you lowered the threshold to get 200 total autoflagged posts, your fp rate would increase, say to 15% of the 200, giving 170 tp + 30 fp + 330 tp not autoflagged + 470 fp not autoflagged.
Now, assume an improvement gives us only 5% FP rate (cut in half) and reduces the total autoflagged posts by about 10% (from 100 to 90). There are still 1000 total posts in the system, so the final numbers are like 86 tp + 4 fp + 414 tp not autoflagged + 496 fp not autoflagged.
But by lowering score threshold we get a larger percentage of the overall total in the system. Say it's now 200 by the time you hit that 10% rate again. So you would have 180 tp + 20 fp + 320 tp not flagged + 480 fp not flagged. The 10% of 200 vs. 10% of 100 is constant, but you've moved from 410 tp going unflagged to 320 tp going unflagged, a net increase of 90 tp flagged, at the cost of 10 more fp flagged (our accuracy ratio).
Obviously the numbers aren't that extreme, but with the 99.75% threshold requirement we're sure of a 399 tp to 1 fp ratio of the increased posts. We don't know exactly what that increase is, but looking historically at only the posts that were already autoflagged it looks like we get more... plus there would be potentially even more than that.
(Up to the limit of posts the system actually scans.) Due to the complexities of the various autoflagging rules (site filters, numbers of reasons) and the varying nature of future spam vs. current reasons/scoring, we can't know exactly. But I'm confident in my position that the number of tp flags autoflagged will increase under a revised scoring system.
... and thus ends my attempt to take over the monologue crown.
