Change to Bayesian only

McAlly · Cadet Joined: May 31, 2004 Posts: 8 Location: Belgium

Just to let you know.
As I read last night about Bayesian filter (A Plan for Spam : http://www.paulgraham.com/spam.html), I decided to use this (learning filter) as my only filter. Except 1 filter to mark every mail to delete ,coming from wanted source, but were I don't have the time to read them at this moment.

I cleared the white and black list and even deleted the learning.dat file (for a fresh start. I've got MWpro since a week, so no big lost)
No new address are added to the white/black list. Also no Spamcop or FA is used.

This was this morning at 8:00 am, now at 5:00 pm, I've got 28 emails
14 spam and 14 mail from Spamcop for confirmation

The last 5 spams were already recognized.

Not bad I guess. Twisted Evil

Ill keep you posted.

McAlly

Al · Posted: Sun Jun 06, 2004 8:58 am Post subject:

I did basically the same thing, some time ago, except that I kept my [Friends] list and one [Good] mail filter.

Baysian hasn't missed a beat, in over a month.

McAlly · Cadet Joined: May 31, 2004 Posts: 8 Location: Belgium

As I suspected, the first good mails I received were false positives, mark as possibly spam.
I send a email to someone i didn't knew his correct address, so I used some 4 different email-addresses and send a copy to myself.
From the mail-server(s) I got 2 mail back saying it was the wrong address and off course my own copy.
All 3 were marked a possibly spam.

MWpro also found a virus (a w32.Beagle.X@mm)

Up till now, I've got 110 email (72 spam)
Bayesian found 41 off them.

If you keep in mind it took 25 mail, before it marked the first spam.

I got 85 email (58 spam) with 41 found (70% found after day +1)

rogerw · Posted: Sun Jun 06, 2004 12:31 pm Post subject:

Bear in mind that several hundred emails need to be used in training for Bayesian to begin to make accurate suggestions.

You'll need to try to make sure that you have roughly an equal number of good and bad emails (or training corpus files of roughly the same size), or the weighting factors for words that good and bad emails have in common might tend to favor whichever type you have the most of.

One suggestion that might help make you happy with Bayesian is to stop training once you're getting acceptable results. Continuously adding to the training will tend to make the decisisions appear to 'drift': as more words/samples are added to the 'bad' corpus - the weighting of those words will change to favor more 'junk' classifications. Likewise with adding to the 'good' corpus.

However, the content of spam might change with time, so training should never be out of the question.

I suggest:


	Home ˇ Topics ˇ Submit News ˇ Top 10 This entire site, Cops Themes, and Computer Cops are Š 2002 - 2004 Computer Cops, LLC. All rights reserved. You can syndicate our news using the file RSS 0.91, ultramode.txt, or RSS 1.0. Acceptable Use Policy. Use signifies your agreement. Engine Copyright Š 2002 by PHP-Nuke, GNU/GPL Licensed. ICRA Member. Paul Laudanski, Member of Computer Cops, LLC Server Load: 888 pages served in previous 5 minutes. Page Rendered: 0.396 seconds.