Timeline for A machine can flag spam automatically. Can it do better?
Current License: CC BY-SA 4.0
116 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 4, 2018 at 16:45 | history | edited | Undo | CC BY-SA 4.0 |
added 625 characters in body
|
| Jul 4, 2018 at 16:35 | history | edited | Undo | CC BY-SA 4.0 |
added 625 characters in body
|
| Mar 19, 2018 at 3:28 | history | edited | Undo | CC BY-SA 3.0 |
edited body
|
| Mar 19, 2018 at 3:22 | history | edited | Undo | CC BY-SA 3.0 |
new numbers
|
| Mar 15, 2018 at 15:21 | comment | added | omikes | fair enough. i'm a bit of a nervous nancy, technologically speaking. | |
| Mar 15, 2018 at 2:30 | comment | added | ArtOfCode | My numbers in that last comment are outdated, actually - we're currently distributing the last third between around 500 users, not 270. That would easily explain the 8 days figure. | |
| Mar 15, 2018 at 2:24 | comment | added | iBug says Reinstate Monica | @oMiKeY We have a lot of new users since this post was submitted. Slightly less than a half of all the automatic flags are distributed among "new users", so it's quite possible that a single user don't get a flag distributed to their account in a few days. | |
| Mar 15, 2018 at 2:20 | comment | added | ArtOfCode | @oMiKeY The system is weighted to use certain accounts more than others - accounts belonging to Charcoal's core contributors are weighted more highly. IOW, 2/3 of flags are distributed between approximately 30 users, and the remaining 1/3 are distributed between the remaining ~270 users. We're also only flagging around 70 posts per day, not 200. So, I'd expect to see 70 flags/day (1/3 of 210 total flags/day) between 270 users. Randomly distributed, that's a flag every 4 days, if you're set up as permissively as possible. And that's still only approximate math, likely overestimated. | |
| Mar 14, 2018 at 15:37 | comment | added | omikes | I thought it was a neat idea and signed up. The system flags an average of 200+ posts a day, using 300+ supposed active users. In eight days, my account did not cast one vote. I would steer clear of this system, which requires full account permission in order to (not) vote. Very shady. | |
| Mar 12, 2018 at 2:38 | history | edited | Undo | CC BY-SA 3.0 |
Update raw data
|
| Mar 12, 2018 at 2:05 | comment | added | iBug says Reinstate Monica | @Undo You forgot to update the raw number data, lololol. | |
| Mar 12, 2018 at 0:54 | history | edited | Undo | CC BY-SA 3.0 |
added 19 characters in body
|
| Mar 10, 2018 at 14:26 | comment | added | J F | @Mazura All of our tests can be found here. We add up the accuracy percentages, then flag based on the total. | |
| Mar 9, 2018 at 0:54 | comment | added | Mazura | Depends on what the criteria is for a 99% confident flag. Is that public knowledge? (It prob shouldn't be ;) If it's people's emails, phone numbers, and general animosity, then bump it up to six and be done with it. Anything else requires human intervention, and SE requires far more than one human to intervene by design. | |
| Mar 8, 2018 at 16:13 | history | edited | Shog9Mod |
edited tags
|
|
| Mar 8, 2018 at 16:11 | answer | added | JosephDoggie | timeline score: -1 | |
| Mar 8, 2018 at 15:19 | comment | added | Hack-R | Seems like this is more of a question for SO users on Kaggle than SO users on SO. But the answer is, yes, of course. What I mean by this is that the spam detection should be made (much) more effective by collaborative and competitive community programming of the detection algorithm as we often do on Kaggle. | |
| Mar 8, 2018 at 11:31 | comment | added | ArtOfCode | @Glen_b I think you may be looking at the wrong boxplot for that one. The all-blue boxplot is the results of last week's experiment. Look instead at the top boxplot, which is of TTD on the top few sites. There's much more data in that one about the TTD for outliers. | |
| Mar 8, 2018 at 1:06 | comment | added | Glen_b | "and a 10x drop in the outliers" -- you have a very skew distribution and only a few dozen points in each group; the number of outliers will itself be a fairly skewed random variable. I don't know that the evidence that there will really be a reduction in the proportion of outliers (by the boxplot rule) is very solid; if you computed some kind of confidence interval for that multiplicative reduction (around your estimate of 10x) I expect it would be pretty wide. | |
| Mar 7, 2018 at 18:18 | history | edited | ArtOfCode | CC BY-SA 3.0 |
no idea how I got those numbers but they're wrong
|
| Mar 7, 2018 at 15:10 | comment | added | Cerbrus | @henning: Consider the time a spam post is "live" on a site. That 400 hours is the total "live" time of all spam posted on a single day. For example: 100 posts in 1 day, with an average time to deletion of 5 minutes would mean 500 minutes in a day. That 400 hours is a order of magnitude more. | |
| Mar 7, 2018 at 14:25 | comment | added | henning no longer feeds AI | What does "400 hours per day" mean? | |
| Mar 7, 2018 at 10:00 | history | edited | ArtOfCode | CC BY-SA 3.0 |
edited body
|
| Mar 7, 2018 at 4:59 | comment | added | jhpratt | @ArtOfCode That's exactly what I'm saying. Don't get me wrong, I don't think it'll happen to anyone, but it is theoretically a possibility. | |
| Mar 7, 2018 at 3:25 | comment | added | ArtOfCode | @jhpratt Not sure I quite see what you're getting at. Are you suggesting that we mark which accounts are used for autoflagging somehow so that SE systems won't automatically flag-ban them if autoflags get declined a lot? If so, that's not necessary - autoflags are accurate enough that they're not going to get declined in anywhere near the volume required for a flag ban. | |
| Mar 7, 2018 at 3:03 | answer | added | Ecnerwal | timeline score: -9 | |
| Mar 6, 2018 at 23:56 | comment | added | Undo | @Ray I went through all the false positives autoflagging has had on Stack Overflow the other day (18). I looked at whether a manual flag had been cast on the post after the autoflags. 2 out of 18 posts had manual flags cast on them - both were far below the threshold we would set for five flags. So I can give you 11% as a very rough starting number. Also, I disagree with the idea that individual flaggers are independent - it would stand to reason that what fools one flagger has a high chance of fooling another. That plays out every day in Stack Overflow's flag queue. | |
| Mar 6, 2018 at 21:56 | comment | added | jhpratt | I would like to see Charcoal setting a boolean on StackExchange's end, such that multiple denied flags in a short time period don't result in a ban for the user. I like the idea, and it has great accuracy, but it's an awfully large risk to take. | |
| Mar 6, 2018 at 21:53 | comment | added | Ray | @Undo I don't have access to the site analytics, so I'm working in the dark here, but what about the figure in Bhargav Rao's answer? It looks like there exist records somewhere of how many spam votes individual answers got even if they weren't marked as spam in the end. That would let us compute the false positive rate per vote rather than per decision. For the bot, those are the same since it either casts all of its votes or none of them. But for humans, the votes are (approximately) independent. | |
| Mar 6, 2018 at 21:44 | comment | added | Undo | @Ray Oh, I mean the data really doesn't exist for posts at this accuracy. There was one false positive and it was handled correctly - but it wouldn't be right to say we had 100% accuracy. You'd have to extrapolate from some other accuracy level. | |
| Mar 6, 2018 at 21:41 | comment | added | Ray | @Undo We don't necessarily need that level of precision for human false positives. If the human false positive rate is, e.g. 2.03%, it isn't a big deal if that gets rounded off to 2%. And we could get that level of precision with much less data. | |
| Mar 6, 2018 at 21:27 | comment | added | Undo | Reputation is already taken into account on the 100.00% number; higher reputation (anything above 1) is a strong indicator of a post not being spam and excludes it from many checks. All told, only 25 posts were autoflagged from users with >= 2 rep, only one being eligible for five flags under this change. | |
| Mar 6, 2018 at 21:25 | comment | added | Undo | @Accumulation As you say, 1 in 30,000 (or even less) is a tiny percentage, so data on human false positives on posts in this level of accuracy just doesn't exist. | |
| Mar 6, 2018 at 21:22 | comment | added | Mithical |
@Acccumulation - How are posts being evaluated spam/not spam for calculating accuracy? That's based on human feedback to the system, and the system calculates the accuracy of each "reason" that caught the post based on the feedback that posts caught by that reason get. So if "URL in title" gets more posts marked as true positives than "bad keyword in body", it will have a higher weight and will be considered more accurate than the other reason.
|
|
| Mar 6, 2018 at 20:58 | comment | added | Acccumulation | How are posts being evaluated spam/not spam for calculating accuracy? And how is 99.99% meaningful? .01% is less than the margin of error. How correlated are human false positives? What impact does reputation have on posts being marked as spam? | |
| Mar 6, 2018 at 18:21 | answer | added | David | timeline score: 6 | |
| Mar 6, 2018 at 16:30 | comment | added | Undo | @simonalexander2005 None that I'm aware of, but it's a valid point. We'd likely see any new system like that come up (or CMs would) and work with them on how to avoid double-flagging a post with similar heuristics. | |
| Mar 6, 2018 at 16:25 | comment | added | Undo | @CodyP There's two things here: Dropping the number of human eyes is a great benefit. Andy is also correct: "we're hoping to see a moderate reduction in the average times, and a significant reduction in the top outliers.". Take a look at that graph - getting rid of the 120+ second outliers is another motivator. There are multiple good outcomes from this, all equally valid (and hard to fit all of them in one meta post short enough for anyone to read). | |
| Mar 6, 2018 at 16:21 | comment | added | RealAnswersNotAI | @Undo In response to your response about how the point is "dropping the number of human eyes that have to look at this stuff from 3 to 1 is a huge benefit IMO". However according to Andy "The major thing we're looking for out of this is a reduction in time to deletion." I understand reducing moderator views from 3 to 1 is a good point, but either we need to critically evaluate the value of 40-seconds-less median TTD, or Andy should change his question to deemphasize TTD. | |
| Mar 6, 2018 at 16:11 | comment | added | simonalexander2005 | Are there multiple spam-flagging systems out there? As in, if Charcoal auto-voted 5 times could another autoflagging system provide the 6th flag? | |
| Mar 6, 2018 at 16:09 | answer | added | WGroleau | timeline score: -4 | |
| Mar 6, 2018 at 15:09 | comment | added | Lightness Races in Orbit | I just want to say, as much as I use Stack Overflow, I never see any spam on it, and that's a real testament to how well the system already works. Good job. | |
| Mar 6, 2018 at 14:37 | review | Suggested edits | |||
| Mar 6, 2018 at 14:57 | |||||
| Mar 6, 2018 at 13:52 | comment | added | Magisch | @Peterverleg Everything we do at charcoal is 100% public record. The project itself is open source and open to all members of stack exchange. Such a thing would get noticed incredibly quickly. So no, we don't have that power. | |
| Mar 6, 2018 at 12:53 | comment | added | Andy | That's not the way the system, current checks and balances, or the project work. We are already running our spam prevention and detection. Everything we've seen and removed it auditable. | |
| Mar 6, 2018 at 9:03 | comment | added | Magisch | @IanKemp Thats a bit of an unfair comparison, if you mean charcoal, that's a arge group effort | |
| Mar 6, 2018 at 8:44 | comment | added | user136089 | @ArtOfCode, re: "nuke", also see Shog9's comment here. And to address a statement you made: I did not say the OP's usage was "invalid". I said it was liable to be misunderstood. (There's a difference.) I also explained why it was likely to be misunderstood. Please accept that feedback in the constructive spirit in which I gave it, and let's avoid outgroup derogation. Thanks :) | |
| Mar 6, 2018 at 8:42 | comment | added | Ian Kemp | Yet another example of why SO should just open-source everything and leave its users to build a better experience than its dev team of (apparently) 1 person. | |
| Mar 6, 2018 at 5:33 | comment | added | ArtOfCode | @Andrew Not quite what you're thinking - there's one post that, had we been running autoflagging at the time, would have been flagged under the settings we're proposing for 5 flags. That's this post (the question marks are standing in for Korean characters that weren't stored correctly as UTF8). | |
| Mar 6, 2018 at 4:45 | answer | added | user310756 | timeline score: 6 | |
| Mar 6, 2018 at 3:19 | comment | added | Andrew | Can u share the 1 post that was incorrectly marked as spam? Curious. | |
| Mar 6, 2018 at 0:29 | comment | added | user136089 | @ArtOfCode, "we've used the word "nuke" here ... in line with how the rest of SE uses it". That is clearly false, or else a longstanding (~9-yr) SE user like me would have understood your usage. Evidently, there is an in-group of SE users that has developed its own jargon, perhaps unconsciously, that is at odds with conventional usage. If I were among that in-group, and hoping to communicate effectively with other SE users, I would be grateful to have been alerted to the likely incomprehensibility of that jargon outside of the in-group. I hope my comments here will be viewed in that spirit. | |
| Mar 6, 2018 at 0:12 | answer | added | Bhargav Rao | timeline score: 42 | |
| Mar 5, 2018 at 23:37 | comment | added | ArtOfCode | @sampablokuper How we've used the word "nuke" here is in line with how the rest of SE uses it. I appreciate it's possible to not have come across that usage before, but that doesn't make it invalid - it simply has a slightly different meaning on SE than you're used to. | |
| Mar 5, 2018 at 23:36 | history | edited | ArtOfCode | CC BY-SA 3.0 |
added 120 characters in body
|
| Mar 5, 2018 at 23:22 | history | edited | ArtOfCode | CC BY-SA 3.0 |
edited body
|
| Mar 5, 2018 at 23:07 | comment | added | user136089 | @Ajedi32, I think that what you are asking for is pretty much what the OP is proposing. Unfortunately, the OP's misuse of the term "nuke" has clouded that fact. | |
| Mar 5, 2018 at 23:04 | comment | added | user136089 | Guide for the perplexed: the OP uses the word "nuke" here against the conventional sense, and in fact means something more like "quarantine". I.e. not irrevocable deletion, merely revocable suppression from the default Q&A listings and search results until further manual review. (This misuse of the word is unfortunate and seems to have already clouded the discussion in several cases. I would be grateful if the OP could edit the post for comprehensibility.) | |
| Mar 5, 2018 at 22:56 | answer | added | Shadow | timeline score: -1 | |
| Mar 5, 2018 at 22:55 | answer | added | Shog9Mod | timeline score: 76 | |
| Mar 5, 2018 at 22:29 | answer | added | DVK | timeline score: 2 | |
| Mar 5, 2018 at 21:06 | comment | added | Undo | @CodyP Ultimately, each spam post takes six flags to delete (ignoring moderators for a moment). Dropping the number of human eyes that have to look at this stuff from 3 to 1 is a huge benefit IMO. Some of it is nasty, nasty stuff designed to haunt the mind - at a minimum, we're cutting the number of people who have to view it to destroy it by that much. | |
| Mar 5, 2018 at 20:36 | history | edited | doppelgreener | CC BY-SA 3.0 |
missed this bit during revision 8
|
| Mar 5, 2018 at 20:01 | comment | added | user291305 | There should be Six Flags Over Georgia. | |
| Mar 5, 2018 at 19:46 | comment | added | Andy | @CodyP View counts are cached for "a bit" - I don't know the exact time period - so accurate numbers aren't possible, at least not via the API. | |
| Mar 5, 2018 at 19:17 | comment | added | RealAnswersNotAI | Do you have any statistics on how many page views happen during that reduction from 50 to 10 second reduction in time? I don't care about 40 seconds difference unless you can prove that a significant number of people see the spam during that time. | |
| Mar 5, 2018 at 19:15 | answer | added | user136089 | timeline score: 18 | |
| Mar 5, 2018 at 18:29 | comment | added | bmike | Could you edit the post - it’s not clear if the time to deletion is measured from the moment the post is live until the post is deleted or the timing is from the first time the system or a person contemplates / takes action on the post. | |
| Mar 5, 2018 at 17:57 | history | edited | Undo | CC BY-SA 3.0 |
Words! https://chat.stackexchange.com/transcript/message/43220299#43220299
|
| Mar 5, 2018 at 17:35 | comment | added | Dessa Simpson | @Undo Alright, not a problem. I've signed up for it, by the way | |
| Mar 5, 2018 at 17:31 | comment | added | Undo | Thanks @DuncanXSimpson; it's a known issue that pops up occasionally. Should be fixed now. | |
| Mar 5, 2018 at 17:28 | comment | added | a spaghetto | @DuncanXSimpson Probably not a bug per se; we were out of disk space for a moment there | |
| Mar 5, 2018 at 17:22 | comment | added | Dessa Simpson | @Undo Absolutely! pastebin.com/EbiJM7WQ Also, I can't reproduce any of the others now that I'm authenticated, so it makes sense that those bugs were missed | |
| Mar 5, 2018 at 17:19 | comment | added | Undo | @DuncanXSimpson Mind throwing me a stack trace in a gist or image? | |
| Mar 5, 2018 at 17:13 | comment | added | Dessa Simpson | Your website at metasmoke.erwaysoftware.com is throwing errors whenever I go to half the pages on it. Doesn't exactly instill confidence. | |
| Mar 5, 2018 at 17:09 | comment | added | Andy | @Ajedi32 SE Dev time is limited, unfortunately. We've had discussions in past about how we can better be integrated. There are also other feature requests for new/improved moderator tooling that would certainly help with some of the other problems mentioned in this post. At the moment, though, asking for large system changes from SE seems to be a non-starter. | |
| Mar 5, 2018 at 17:06 | comment | added | Ajedi32 | As an alternate solution to this problem, could StackExchange implement a system that hides posts with multiple spam flags from the site until they are dealt with in a review queue (one way or the other)? That would accomplish "You will see a reduction in the time spam spends on the site before being deleted." and "Fewer humans will have to involuntarily see each spam post." without reducing human oversight on the flags. | |
| Mar 5, 2018 at 16:50 | comment | added | Glorfindel Mod | @ChrisW questions: 28036 vs. 2470 answers. | |
| Mar 5, 2018 at 16:47 | comment | added | TheLethalCarrot | @Catija Aye but I have a few accounts but I'm only active on SFF. My SO and PPCG accounts are both >5k rep but I'm not active on any of them. Having the account doesn't mean you're active on it. I do see your point though. | |
| Mar 5, 2018 at 16:44 | comment | added | Catija | @TheLethalCoder If you don't have an account on a site, you can't autoflag there... so unless you've joined all of the sites, you're not going to have to limit it too much. :) | |
| Mar 5, 2018 at 16:41 | comment | added | ChrisW | Is it mostly answers that you end up flagging, or mostly questions? | |
| Mar 5, 2018 at 16:35 | comment | added | TheLethalCarrot | FWIW I support this proposal and would like to see it go forward, I'd just like more community input on a per site basis before applying it network wide. | |
| Mar 5, 2018 at 16:30 | comment | added | TheLethalCarrot |
@Andy Ahh okay, it was this line that got me thinking about that who may or may not be active on the site. And I'm not really sure I like someone not familiar with a site auto flagging content because that makes them harder to get in contact with in case of any mistakes. Even though they can likely be contacted through other means.
|
|
| Mar 5, 2018 at 16:28 | comment | added | TheLethalCarrot | Also as others have said I like the idea of setting this up per site with a discussion on that site beforehand. On SO it's very beneficial and probably worthwhile having more auto flags. Whereas, on SFF we already clean up quite quickly. It'd be nice to change settings (On/Off, No of auto flags) per site depending on that sites community consensus. | |
| Mar 5, 2018 at 16:28 | comment | added | Andy | @TheLethalCoder You can already select which sites you want to flag on. | |
| Mar 5, 2018 at 16:26 | comment | added | TheLethalCarrot | One thing that I'd prefer to change is users who opt in choose what sites they can auto flag on. I'm not active on SU and I wouldn't really feel comfortable flagging over there even if it is for the benefit of the site. However, I am active on SFF and would be comfortable letting my account auto flag. Would this be something you'd consider adding/changing? | |
| Mar 5, 2018 at 15:14 | history | edited | Undo | CC BY-SA 3.0 |
added 59 characters in body
|
| Mar 5, 2018 at 15:06 | history | edited | TarynStaffMod |
edited tags
|
|
| Mar 5, 2018 at 14:16 | answer | added | Monica Cellio | timeline score: 70 | |
| Mar 5, 2018 at 13:25 | comment | added | ArtOfCode | We'd like to, @JoshCaswell, but that requires integration with SE and SE dev time, which is short right now. | |
| Mar 5, 2018 at 13:22 | comment | added | jscs | Sounds like this system should be put on the front end, checking posts before they are even published, instead of having to wait for individual users to raise flags (even automatically). | |
| Mar 5, 2018 at 12:30 | history | edited | ArtOfCode | CC BY-SA 3.0 |
don't break
|
| Mar 5, 2018 at 12:26 | history | edited | iBug says Reinstate Monica | CC BY-SA 3.0 |
added 6 characters in body
|
| Mar 5, 2018 at 12:19 | history | rollback | ArtOfCode |
Rollback to Revision 8
|
|
| Mar 5, 2018 at 12:11 | history | edited | iBug says Reinstate Monica | CC BY-SA 3.0 |
deleted 58 characters in body
|
| Mar 5, 2018 at 10:59 | history | edited | ArtOfCode | CC BY-SA 3.0 |
literally just changed
|
| Mar 5, 2018 at 10:30 | history | edited | ArtOfCode | CC BY-SA 3.0 |
added 10 characters in body
|
| Mar 5, 2018 at 10:26 | history | edited | doppelgreener | CC BY-SA 3.0 |
made this image a link to the full size graph, since the text is illegible on here
|
| Mar 5, 2018 at 9:33 | answer | added | Petter Friberg | timeline score: 35 | |
| Mar 5, 2018 at 8:58 | answer | added | Mad Scientist | timeline score: 126 | |
| Mar 5, 2018 at 5:19 | comment | added | Undo | @Ano There was a big wave on Workplace and ELU a while back; three flags was meant to be a stopgap and just... never got changed. It's not really worth discussing in depth here; those sites are a small portion of the network in any statistic you look at. | |
| Mar 5, 2018 at 5:16 | answer | added | Journeyman Geek | timeline score: 28 | |
| Mar 5, 2018 at 5:16 | comment | added | Sonic the Anonymous Hedgehog | @Shog9 Did those two sites get their thresholds changed before Smokey started to take off? Is it because of Smokey that there are plans to get them changed back to six flags? | |
| Mar 5, 2018 at 5:14 | comment | added | Shog9 Mod | There are only two "three flag sites" and there will probably be zero in the near future. Let's not get sidetracked. | |
| Mar 5, 2018 at 5:13 | comment | added | Sonic the Anonymous Hedgehog | @Undo This was not clear from the original post. Also, I'd recommend editing in what is the plan for those 3-flag sites. | |
| Mar 5, 2018 at 5:12 | comment | added | Undo | @Ano 'Excluded from the experiment' means 'we kept the status quo' - the status quo for the 3-flag sites is currently a single flag. | |
| Mar 5, 2018 at 5:11 | comment | added | Sonic the Anonymous Hedgehog | @Undo ArtOfCode says they get just one autoflag, but you say they get none...who's correct? | |
| Mar 5, 2018 at 4:53 | comment | added | Undo | @Ano They're excluded from the experiment above. If we go to 5, I would expect those two sites to go to 2 flags (one manual to nuke, just like the 5 on a 6-flag site). | |
| Mar 5, 2018 at 4:52 | history | rollback | ArtOfCode |
Rollback to Revision 3
|
|
| Mar 5, 2018 at 4:52 | comment | added | Sonic the Anonymous Hedgehog | What about sites that only require three flags to nuke a post? Will that be increased to two autoflags? | |
| Mar 5, 2018 at 4:51 | history | edited | Sonic the Anonymous Hedgehog | CC BY-SA 3.0 |
spelling and grammar; make it clear from the very beginning that this is done on behalf of willing users, if I were an unfamiliar user reading this, my first instinct is that you are using a bunch of bot accounts
|
| Mar 5, 2018 at 4:38 | comment | added | ArtOfCode | Just for you, @BhargavRao. | |
| Mar 5, 2018 at 4:37 | history | edited | ArtOfCode | CC BY-SA 3.0 |
Definitely 5 we're proposing
|
| Mar 5, 2018 at 4:37 | comment | added | Bhargav Rao | Duh! still no free hand circles? Where do I look? | |
| Mar 5, 2018 at 4:35 | history | edited | ArtOfCode | CC BY-SA 3.0 |
Definitely didn't miss taking that out from an earlier revision
|
| Mar 5, 2018 at 4:34 | history | asked | Andy | CC BY-SA 3.0 |
