On Fri, August 9, 2013 1:01 pm, RW wrote:
> BAYES works on rendered text it doesn't see the HTML.
Hmmm. It doesn't see HTML comments, which would appear in rendered HTML
source even though they are "invisible?" OK, in that case, I have NO idea
why the spam isn't hitting Bayes, because it looks pretty damn spammy to
me. I wonder if it's the heavy use of images, but I don't know.
> Do you actually get a significant amount of ham between 0.99 and 0.999?
> Personally I only get 1 in 1000 above 0.55, and nothing above 0.65.
Ham, absolutely not. So yes, I suppose I could just treat all Bayes99 as
if it were Bayes999 and score it more highly than I do. Right now I have
Bayes99 at 4, Bayes999 at 4.5. I could eliminate Bayes999 and make
Bayes99 score 4.5... but I do worry a little bit about FPs, even though I
guess I shoudn't, statistically speaking.
On the other hand, one could consider making Bayes999 a poison pill.
Generally spam will only rank there if you've learned something nearly
identical to it. At that point, perhaps it might be worth just scoring it
with 5 or higher (assuming your threshold is 5, as mine is).
--- Amir
