Thomas Barth <tba...@txbweb.de> writes:

> Am 2024-05-11 21:54, schrieb Bill Cole:
>> I have no idea who the Debian "spam analysts" are but I am certain
>> that they are not doing any sort of data-driven dynamic adjustments
>> of scores based on a threshold of 6.3 nor are they (obviously)
>> adjusting that threshold daily based on current scores.
>
> I found the passage in my old Postfix book. The author writes: "It is
> recommended not to carelessly set the value of $sa_kill_level_deflt to
> any fantasy values. The score of 6.31 is not arbitrarily chosen, but
> the statistically calculated optimum for the best possible spam filter
> rate with as few false positives as possible. If you increase the
> value, more spam will get through; if you lower it, your false
> positives will increase."

The comments about adjustments are true, but the idea that it is optimum
is flat-out nonsensical.

The key question is how you weight a false positive compared to a false
negative.  Only after you decided that can you pick an optimium, for a
given corpus of already-received mail.

> It may be that the value is outdated, but that is for the maintainers
> of the relevant Debian package to decide. I'll just adapt my rules to
> this one value.

That is an odd position.  It is very easy to set the threshold in a
local config.   Deciding instead to adjust scores to an oddball
threshold seems bizarre to me.

Personally, I don't use the 5, but instead have shades of grey, where
>=1 is binned into mailboxes that are "maybe spam" through "very likely"
spam, and at some score, I reject at the MTA level.

I find that legit mail shows up in e.g. spam.2 (>= 2 and < 3), but it is
almost never mail that I would be upset to have missed (but I don't) or
mail that I would be upset to not get in a timely manner (I only see it
every day or so).  However, this really drops the FN rate of spam in my
INBOX, which matters a lot to me.    Basically I consider a FP into my
"spam.1" mailbox, as long as it isn't really important to me, to be not
a big deal at all, and I'd rather have 10 or those than 1 FN in my
INBOX.  But, actually MTA-rejecting mail that I shouldn't, a FP at that
level, is a big deal, and I avoid it.  I think it's about one message a
year -- and while it's ham, it's very spammy ham.

Reply via email to