[Spambayes] Training good messages has no effect
Tim Peters
tim.one at comcast.net
Tue Aug 12 00:40:54 EDT 2003
[Hannes Scheidler, having troubles]
I strongly second Mark's suggestion that you try reducing your 100::1
ham::spam training-data ratio to something closer to 1::1.
Another thing you might want to try, based on the nature of your ham, is to
look for your spambayes .ini file and change this line:
replace_nonascii_chars: True
to:
replace_nonascii_chars: False
That isn't so important, though -- it probably won't help much. If you keep
your 100::1 ratio and change the line:
experimental_ham_spam_imbalance_adjustment: True
to:
experimental_ham_spam_imbalance_adjustment: False
then I'm pretty sure you won't see these msgs classified as spam anymore --
but you may not see *any* messages classified as spam anymore then. It's
really not known what will happen with such a large imbalance in training
set sizes (the code was never tested that way).
More information about the Spambayes
mailing list