http://plg.uwaterloo.ca/~gvcormac/spamcormack1.pdf
This study is much better than most analysis of different spam filters,
in that it deals with SpamAssassin in different modes of Bayesian
learning, and cares about false positives as well as false negatives.
However, it's needlessly obscure in it's terminology and statistics.
The critical summary info is on page 14 in tables VIII (ham
misclassification, by which he means false positives) & IX (spam
misclassification, by which he means false negatives).
Not surprisingly, SpamAssassin Bayes supervised (i.e., mistake-based
training) works best.
- dan
--
Dan Kohn <mailto:[EMAIL PROTECTED]>
<http://www.dankohn.com/> <tel:+1-650-327-2600>
