Am 21.09.2016 um 18:28 schrieb Thomas Barth:
Am 21.09.2016 um 18:00 schrieb li...@rhsoft.net:
the problem of the OP is that he starts things the other side round and
first reject without good evidence and don't have anything to make the
system bullet profe because it's rejected
I remembered that I read a book about Postfix with the topic "Training
with SpamAssassin". And the author was against additional training. The
more you train the worst the result. With the motto "I cook an egg for
more than 15 minutes, but it is still hard." They re other arguments for
not autolearning, but my english is not that good to translate a
complete chapter. And if there are some mails breaking through the wall,
than it is better to create rules against the header. Clear facts
without side effects.
He also wrote that Amavis/SpamAssassin is learning itself. Each mail
classified as spam with a score of more than 12.0 is learned as spam and
there should be a logfile entry with loglevel 2 if a mail has been
learned as spam. I never increased the loglevel to check that.
I followed his opinion because it is the best book I ve got
(www.postfix.de, next SpamAssassin/Amavis training course in November, I
m thinking of participation)
"against additional training" and "other arguments for not autolearning"
are the exactly *opposite*, however, i can assure you that a well
trained bayes with any autolearning reachs a 90-95% hit quote proven by
5 false positives and 30 spamreports on some hundret users in 2016
autolearning is anyways bad because it tends to classify alread FN oder
FP in the exatcly wrong direction - you need to train *wrong classified*
mail where you are 100% sure if it's spam or ham and just ignore
anything where you are unsure, the rest will have common patterns which
are learned over time with your well classified ones
anyways, a spamfilter completly without bayes and URIBL not wroking has
no business to run in production
