On 20 Dec 2006, pinard@iro.umontreal.ca wrote:
One (relatively) recent problems with Bayesian filters is that many SPAM messages are constructed in such a way that they poison your filters. If you train such messages as SPAM, they contain enough "good" text, which is irrelevant to the SPAM itself, so to progressively weaken the discriminatory capabilities of your filters. The effect shows in the long run. It's a tad sad, because Bayesian were once highly effective.
Yes, unfortunately combating spam is a full time job for the Spam filter authors. I even have URIDNSBL and razor2 turned on in my Spamassassin install, but it is still not 100%. Charles -- DPRINTK("Last time you were disconnected, how about now?\n"); linux-2.6.6/drivers/net/tokenring/ibmtr.c