Okay. I've just done some data flow modelling of how this autoflagging system needs to work, and there are things we need to expand on before we can even start building the platform for it. I'm gonna create a google doc detailing those, with a RFC in it for everyone to pile in on.
I got interesting results of trying to run a neural network using the existing smoke detector reasons to flag posts (so the neural network never looked at the existing feedback). Running on the last 10 000 posts (8 110 had valid feedback of those), it found a pattern to flag 105 posts of those, with a helpful rating of 100% (in simulation based on the feedback given to MS)
@Undo reasons as bitmap & a bit if it was a answer or question
The network has an input layer with 73 neurons (all active reasons in the last 8 110 valid posts), a hidden layer with 7 neurons, and an output layer with 1 neuron
@Ferrybig question: we know that probably around 80% of those 10k posts were actually spam. This thing found what, 105 posts to flag? Is there some way you can increase the number of posts it finds to flag and still keep it within acceptable bounds? There's a lot more there that it could flag
Another question: what controls do users need to have in this system? Obviously, we want (a) use or don't-use my account for flagging, (b) a limit on numbers of flags per day. Anything else?
It's actually not too difficult technically. Add a weight field to the Reason model, and a scheduled task once per day to update the weights automatically.
@ArtOfCode quick status update, its now flagging around 5314 posts (out of the 8800 ish it know to have valid feedback) and is getting a helpful percentage of around 78, it will sort out the best ones at the end, and continue with them
@Ferrybig oh, that's something I hadn't thought of
though we can get around that with this structure by not creating the link record if the user is a moderator, so that the user isn't included in FlagSite.users for a diamond site
Aye, but I'm trying to work out the calls we need to find what posts it would catch from that giant repository, and if there's enough and if it's accurate enough.
When you say "I wanna automatically flag things on this site matching these conditions", we check the database for what that would have caught in the last <timeframe>. If it matches some sanity checks we have, you can create it. If it doesn't, then we just don't allow creating of it
@ArtOfCode We can figure that one out easy enough, I'm not worried about it
Honestly, I think I'd do that by working out what sane values are, and just validating to restrict to those. As in, you can choose the max rep you want to flag, but it can never be over 100. You can choose your minimum weight, as long as it's over 5. Etc.
I've got a list of users who could flag based on the conditions set. Now what we need is to select users out of that list based on who has flags left that they'll allow us to use.
@dorukayhan that's what someone was doing the past couple days, on Ask Ubuntu, got squished with bans finally, but it prompted this blacklisting. Tried to disguise the spam link as a link to "software" that solves the original question...
@Undo @ArtOfCode wrt the autoflagging system, should we be concerned about 'level of information accessible' by the autoflag system admins? If we are, then should we be looking at SE Mod Agreement as a limiting factor for admins (so that there's the implied agreement for admins to properly handle PII when it comes across their plate)
another one close to spam... note OP has two identical answer stackoverflow.com/a/41152851/5292302, however I can't judge if app has any sense related to questions