Could our spam filter be made extraordinarily paranoid about answers with emails in them?

Question

RPG.SE has been receiving spellcaster spam for years (such as in this 10k only example). I understand that there's some machine learning behind our spam flags, but this stuff seems to regularly fall through the cracks. That's probably because we talk about priests and spellcasting and witch doctors all the time anyway, and thus train the machine that this stuff is fine. Its natural learning seems to not be adequate in this case, and I think it needs some manual guidance to deal with this problem.

This spam has other factors in common beside its subject matter we can use to identify it: it always arrives as answers, it usually contains no formatting or line breaks (proper, double-enter ones - single-enters seem common), probably contains a lot of capitalised words, and they always contain an email address.

I got curious how many of our valid posts actually contain emails - we have only two. Here's a data explorer query that makes a naive search for stuff that might even vaguely resemble an email address (thanks to Shog9). Of the 19 results, most are self-censored expletives, @-mentions, and anydice syntax. Putting those aside, here's the two that actually do contain email addresses:

• What is an "Irensavelite"
• What campaign wiki - or other collaborative tool - has these features?

Both of these are questions.

So, we get spellcasting spam answers containing email addresses regularly, and we have absolutely no valid answers that contain email addresses (we may have received ones that have since been edited, like those people who say "hey contact me at [email protected] if you want to know more").

Could our spam filter be tuned to be super paranoid about answers with email addresses? Like, as in Matrix anyone-could-be-an-agent levels of paranoia? Especially when the post contains no other formatting or double line breaks.

Answer 1

This is the problem - we can filter it, but they quickly just obfuscate the email address using some other means. Text filtering does not work, which is why the spam system operates on different levels, we can't just go by the text being submitted.

However, I'm going to be modifying the weights that this site carries when it comes to teaching the system, the following actions from users and moderators here will cause it to learn more rapidly:

• Spam-flag deleting a post
• A moderator destroying an account for spamming
• Tripping of various 'honeypots' that we have laying around, because we like honey

Basically, when this community identifies something as spam, the system will acknowledge and deal a blocking blow sooner than it would otherwise. This 'spellcaster' stuff has been appearing all over the network for quite a while, but I do agree that it's a bit more concentrated here due to the topic.

But please, don't spend too much time considering ad-hoc filters to try to keep certain things out; they quickly learn and just work around it - it ends up just being a waste of time and resources.