On 12/18/2009 06:34 PM, John Hardin wrote:
T_RCVD_IN_RP_CERTIFIED
SPAM% 0.0851 126 of 148025 messages HAM% 0.3738 746 of 199558 messages
S/O 0.185 RANK 0.63
Frack.
T_RCVD_IN_RP_SAFE SPAM% 0.0851 126 of 148025 messages HAM% 2.1367 4264
of 199558 messages
S/O 0.038 RANK 0.80
The weekly corpus is too small to safely draw any conclusions.
https://issues.apache.org/SpamAssassin/show_bug.cgi?id=6247#c54
The rescore masscheck however has indicated that whitelists do more harm
than good to spamassassin's classification of ham vs. spam. These weeks
of whitelist flamewars were hot in emotions but lacking statistical
data. The numbers suggest that the whitelists are benefiting the
whitelist purveyor, while having zero or minutely negative impact on
spamassassin.
If this is indeed true, then we might want to ask ourselves why we are
giving such influential weights to the whitelists. Even the newly
reduced scores like -2 for RP_SAFE may be unearned given the statistics
we are seeing. If we had left these rules floating, the GA rescorer
would have given much lower scores.
I might be in favor of further score reductions in the short-term until
we figure out a long-term plan.
In related news...
DNSWL
To my surprise, Matthias has begun to implement my recommendations of
improved manual abuse reporting, and automated abuse reporting. Their
accuracy even without automated abuse detection isn't too bad.
ReturnPath
If their self-description of abuse detection methodologies is indeed
true, then I don't know what ReturnPath can do more to improve.
score RCVD_IN_IADB_VOUCHED 0 -2.2 0 -2.2
score RCVD_IN_IADB_DOPTIN 0 -4 0 -4
score RCVD_IN_IADB_ML_DOPTIN 0 -6 0 -6
Any idea what this is? It seems these aren't hitting anything in the
weekly masscheck.
Warren Togami
wtog...@redhat.com
