Frank Solensky <fsolensky(_at_)premonitia(_dot_)com> writes:
Just posted on slashdot: a Bayesian approach to the problem that reports
to have rates of 0.5% on false positives and 0% false negative:
http://www.paulgraham.com/spam.html
Nice short-term approach.
Unfortunately, easily defeated with just appending (perhaps as an HTML
comment) a long innocent-looking fragment (e.g., a 30KB piece from a
random book).
Further, in its *present* form, where unfamiliar words are given 0.2
spam probability, easily defeated by just adding a lot of randomly
generated `words' like 9nscS9Ft, iuiF0kKw, 6AycPEbU, nsUdjGeP, etc.
Given enough of these, the Bayesian probability formula will declare
even a piece of mail that consists of a sales pitch for a pornographic
web site have a probability of being spam that is arbitrarily close to
0.2.
--
Stanislav Shalunov http://www.internet2.edu/~shalunov/
"Which one is worse? Both are worse." -- V. I. Lenin