> Bad keyword in body, bad keyword in title, bad keyword in username ---------- Title - Position 1-14: design diagra Body - Position 1-14: design diagra Username - Position 1-14: design diagra
this grappling with why the ever-increasingly complex regexes are producing false positives I think is one of the reasons to look into automating the filtering with machine learning
[ SmokeDetector | MS ] URL-only title, bad keyword in body, bad keyword in title, link at end of body, pattern-matching website in body, pattern-matching website in title, repeated URL at end of long post: maleenhancementshop.info/hyper-nitro-core/ by Hyper Nitro Core on askubuntu.com (@ThomasWard)
[ SmokeDetector | MS ] URL in title, link at end of body, pattern-matching website in body, pattern-matching website in title: healthpluscogni.com/aviqua/ by sajoso on drupal.stackexchange.com
'Need to get 235 mB worth of archives. Do you want to proceed? (y). Downloading 84 packages at 100 kB/s' =(. I forgot how slow the wifi dongle is on my Pi
@angussidney I wanted to be smart with my pi, and I used tmux to update large amounts of packages overnight, but the next day I couldn't connect back because tmux was updated and didn't want to connect to the outdated server running the update
Well, plenty is a gross exaggeration. Still, I'm not convinced that user is associated with that specific website. He was probably just looking to score some quick points.
[ SmokeDetector | MS ] Bad keyword in link text in body, bad keyword in title, bad keyword with email in body, blacklisted website in body, repeated URL at end of long post: TERADATA ONLINE TRAINING IN USA by elvaanan5511 on stackoverflow.com (@dorukayhan)
@tripleee you're not, I guarantee it. Quotas are done on a per key and per IP basis - even if you use smokey's key, you're on a different IP so it won't affect the quota that Smokey uses
@ArtOfCode if I'm reading the SDML stuff correctly, the primary problem was the lack of samples, and that it wasn't trained on non-spam samples at all AFAICT
but I'm trying a multinomial approach which should hopefully have higher precision
I want to collect samples over a slightly longer period of time to avoid introducing any bias from this particular point in time, but that means it will take some time to get it up and running even for a quick proof of concept
@ArtOfCode yeah but there's significant bias in the FPs, you want pure clean samples, and much more than the trickle that is FPs
the "normal" training set should be as normal as possible
so I'm simply excluding any NAA or FP answers from MS
and was thinking of requiring upvotes on the training samples down the road
