On Monday 17 November 2003 08:16 am, Michael V. Sokolov wrote: > I've noticed message with such headers: > > X-Spam-Flag: YES > X-Spam-Checker-Version: SpamAssassin 2.60 (1.212-2003-09-23-exp) on > antispam X-Spam-Level: ****** > X-Spam-Status: Yes, hits=6.0 required=5.0 tests=BAYES_99 autolearn=ham > version=2.60 > X-Spam-Report: > * 6.0 BAYES_99 BODY: Bayesian spam probability is 99 to 100% > * [score: 0.9974] > > But, autolearn=ham by default is set to 0.1 (and it is, yes)? So why this > message was autolearned as ham?
This happens because auto-learning uses (if I recall correctly) ruleset 0, which means no Bayes or network tests are considered in determining whether the message is spam/ham. In my opinion this is an annoying flaw in 2.6 as any spam that doesn't hit any of the standard rules is learned as ham, if the default auto-learn thresholds are used. This leads to a skewed bayes database until you get around to doing an sa-learn --forget and sa-learn --spam on the incorrectly learned messages, and you can only do that if you have access to all learned messages that pass thru, which isn't practical, or ethical, for a site wide bayes. My "solution" to this issue, which isn't ideal but "it works for me"(tm), is to set bayes_auto_learn_threshold_nonspam to a small negative number in local.cf. What that effectively does is make auto-learn never learns ham so you have to feed it ham with sa-learn on a regular basis. The drawback is the database can be poisoned easier torwards scoring everything as spam and tends to score slightly higher on tokens that aren't all that spammy until you do another sa-learn run. In my opinion I'd rather see a false positive as a result of a more spam weighted database rather than a very easily poisoned torwards ham database that's effectively not unpoisonable by me. ------------------------------------------------------- This SF. Net email is sponsored by: GoToMyPC GoToMyPC is the fast, easy and secure way to access your computer from any Web browser or wireless device. Click here to Try it Free! https://www.gotomypc.com/tr/OSDN/AW/Q4_2003/t/g22lp?Target=mm/g22lp.tmpl _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
