Re: Problems with Bayesian filtering

Note that the spammers are now wise to those of us who look for x-on-x 
(eg white-on-white) text.  They have started to play with luminosity, 
so that the text is eg 0xFFFFFE on white...  You'll want to calculate a 
luminosity difference between the ink color and background color rather 
than just compare them for equality.

On ignoring big blocks of "fake" text, you can work around this as well 
by essentially doing noise reduction on the text...  If you see too 
many unusual tokens in a row, you basically ignore those tokens until 
you start seeing sequences of normal tokens again.  The dobly thing in 
the /. article yesterday effectively does something like this.

C

On Feb 25, 2004, at 5:52 AM, <[email protected]> wrote:

> In my implementation of Bayesian mail classification (POPFile) I
> specifically look for the use of trickery like white on white (I call 
> it
> Invisible Ink), 1pt fonts etc. and use them as additional tokens for 
> the
> classifier.   They are in fact very good indicators of spam.