On 7/11/2014 12:15 PM, Angus McIntyre wrote: > Eric Shubert wrote: >> P.S. Anyone else using DSPAM? > I set up DSPAM and have it running, but I haven't had much luck with it. > > I have a vague memory of seeing someone claim that you could train DSPAM > using only 'spam' samples (no 'ham'), so I fed it my entire spam archive > for the month, with only a very small number of 'ham' examples. > Afterwards, DSPAM decided that all email was spam. (Granted, that probably > gives it a false-positive rate of about 3%, but still). I train only on spam, not ham. Also, I do not train corpus.
When I trained DSPAM with the '--source=corpus' and '--source=error' flags it did not work well. When I trained DSPAM only using the '--source=error' flag, messages with a DSPAM signature, this is when I started hitting 97%-99% accuracy. Here is my training procedure: 1) Set up DSPAM to put signature in messages in .qmail-default. 2) IMAP users move spam messages to a spam folder (Outlook 2012 users must use a different email client since it strips lines from email headers). 3) Train only those spam messages that have a DSPAM signature and have not been tagged by DSPAM as spam. DSPAM will ignore messages it has already learned. Do not train messages that don't have any DSPAM signature. Here is may training mechanism which strips SA's SPAM tag (I don't remember why I did this) : cat $email-file | sed 's/\*\*\*SPAM\*\*\*\[sa\]//g' | sed 's/\*\*\*SPAM\*\*\*//g' | dspamc --user $USER@$DOMAIN --mode=teft --class=spam --source=error 4) Set up maildrop and .mailfilter file (called from user's .qmail file) to drop marked spam (by DSPAM) into the spam folder. When set up this way a cell phone's inbox will not be cluttered with tagged spam. > > I then blew away its databases and tried to train it more selectively, > letting it rate everything and forcing it to re-learn (as Spam) any > messages that it had wrongly classified as Innocent. Unfortunately, I > never seemed to be able to even "move the needle": feeding numerous > samples of a highly-recognizable type of spam only reduced DSPAM's > confidence that they were Innocent, but it never seemed to tip over into > considering them Spam. I think the best that I ever got was a verdict that > some particularly flagrant piece of spam was only Innocent with 60% > probability. > > I should work out a more intelligent training regime and see if I can't > get it to behave as advertised, but for me it certainly hasn't been the > 'out-of-the-box' spam-recognizing miracle that it has been for some I had to clean individual users' databases a couple times before I got it to work they way I wanted. And, you're right this is not an out-of-box drop-in. I did not expect it to be so I had no disappointment. ;-) > people. > > Angus > > > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
