[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: bayesian filter training question

On torsdag 29 september 2005, 21:51, Roberto C. Sanchez wrote:
> So, I finally decided to get with the 20th century and install
> spamassassin (acutally spampd hooked through postfix) to do site-wide
> spam filtering for my server. 


> My question is this.  As I am training 
> it with sa-learn, is it (good|bad|indifferent) to train it on spam
> that has already been flagged as spam.  That is, will this reinforce
> spamassassin's notion of spam or ruin it?

No, that's fine. In fact, SA has this autowhitelist concept that does 
exactly that (it's not really a whitelist, though, more an "evening out 
weird things that may happen", I'm not using it). 

You should have a good look at bayes_ignore_header, so that it won't 
train on things that are obviously in spam. SA is pretty good it this 
itself, but if you see spam that has been filtered elsewhere a lot, be 
sure to use it.

I'm guessing that you, like me, are doing this for your family. In that 
case, I have found that it is quite sufficient to train a single 
database with the spam and ham of the entire family. If you have more 
diverse users, you would probably need to have a per-user 
configuration. For example, a friend of mine has an uncle who is a 
psychiatrist working with people with gambling obsessions, and SA was 
pretty catastrophic for him until he got a per-user config.

Finally, I found that SA, in it's default 3.0-form was much too 
conservative about the assigned scores, so I have a bunch of rules that 
I have adjusted the score of. You'll get some experience about that in 
time, I guess. Also note that SA 3.1 has been released upstream.


Kjetil Kjernsmo
Programmer / Astrophysicist / Ski-orienteer / Orienteer / Mountaineer
Homepage: http://www.kjetil.kjernsmo.net/     OpenPGP KeyID: 6A6A0BBC

Reply to: