On Wed, Jun 24, 2009 at 09:03:59AM +0200, Patrik Jansson wrote: > Hi, > I've been experiencing a very low accuracy lately. Performance shows > an overall accuracy at 79,570%, which is kind of bad for dspam I > guess. > > I'm always retraining those mails coming through but it doesn't get > any better. What I've noticed is that like 9 out of 10 spams coming > through are sent with a FROM address the same as any of my two > addresses used in this mailbox. > > These are the two most recent spams that have passed the filter: > > Type Retrain Day/Time From Subject > Additional Info > Miss Retrained (Undo) Wed 8:35a "Claris Ybql" > <[email protected]> Someone in your car Delivered > Miss Retrained (Undo) Wed 1:50a "Benton Ygud" > <[email protected]> News on warranty works > Delivered > > (my real addresses have been altered above) > > It seems that dspam is having a hard time figuring out whether or not > the mail is a spam when the FROM address field includes my own > address. Sometimes the mail gets caught but far from often. > > What can be done to fix this? > Thanks, > -Patrik Jansson >
Hi Patrik, What are your dspam_stats? We have seen behavior like this in two cases. In the first, training was done in a fashion where similar messages to an impartial observer were trained as both spam and innocent. The resulting schizophrenia gave poor results. In the second, an inappropriate training mode was chosen relative to the spam/not-spam message ratio. The usual culprit was using TEFT with a very high spam to good mail ratio. This resulted in a disproportionate amount of spam being trained versus the good message. This again reduced the accuracy of DSPAM (That's how Bayesian statistics work.) In both cases, accuracy was restored by reinitializing the corpus in question using an appropriate training process and training mode appropriate for the E-mail in question. Good luck, Ken ------------------------------------------------------------------------------ _______________________________________________ Dspam-user mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/dspam-user
