On 2/17/2014 3:37 PM, RW wrote:
On Mon, 17 Feb 2014 10:21:21 -0500
Kevin A. McGrail wrote:
In the end, all we are doing is adding a gradient for 99.9 to 100%. I
will change this in the base rules for better clarity instead because
we already have evidence it's a good move.
Score often don't follow the strength of a rule seen in isolation. IIRC
when the scores were auto-generated they were pretty flat from BAYES_80
upwards, with BAYES_99 actually being a little lower than BAYES_95.
AFAIK this whole idea of having a gradient just seems to have been
made-up.
You are entitled to your opinion but your email is a bit acerbic. Please
reminder, I am a volunteer on this project and I've given a lot of time
to help block spammers. And I invite you to do the same because less
rules/code/documentation/wiki/advice, etc. more than debate in my
experience.
Beyond that, fighting spam is both an art and a science. Far too much
of what we have to do is identify patterns, throw things to the wall,
see what sticks and then tweak often starting with Art and refining with
Science. In the case of BAYES_999, my expectation on this rule was to
do an analysis on the validity of the change using our RuleQA system but
the rule got auto-promoted in a way that was not intended.
Therefore, I suggest you score BAYES_99 and BAYES_999 the same. This
will make the gradient irrelevant in your setup with an infinitesimal
impact on the system. However, I've been convinced that there is merit
to the change or I wouldn't have bothered.
What is your specific concern with the gradients because all it does is
add flexibility that some users are adamant will help combat spam?
regards,
KAM
