On 01 Jul 2020, at 14:20, Aner Perez <a...@ncstech.com> wrote:
> we have the spam threshold set very low (2.4)
This is a terrible idea and exposes a fundamental misunderstanding of how SA
works.
If SA scores an email as 3.3 then the message is not considered spam by SA. If
you ignore this and mark it as sam anyway, you have no one to blame but
yourself. Reducing the threshold increases the number of non-spam messages that
are marked as spam. It will also have very little effect on actual spam
messages. The only exception to this is if you have a badly trained Bayes, as
that can swing the scoring quite a lot.
Set your threshold back to 5.0 and train your Bayes with actual spam you
receive and actual ham you receive. The best Spam to train is spam that is not
tagged by SA as spam (ignoring the bayes portion of a score). So, a message
marked at 5.5 with BAYES_50 is a price candidate for training as it would be
marked 4.7 without the BAYES_50.
It would have been better, I think, had SA designed the system to score
anything over 0 as spam and anything under 0 as ham as I suspect very few
people would make this mistake, but it's a bit late for that now.
Just think of it this way, when you set the threshold below 5, you are saying
to SA "please mark legitimate mail theat I want to receive as spam."
--
'Oh, them as makes the endings don't get them,' said Granny.
--Maskerade
