That's why i created 2 accounts. spambox & hambox.
when an user checks his webmail, he can click on a button which send the mail as spam in the spambox, or send it as ham in the hambox.
Therefore, i have a new corpus each day, composed with spam & ham messages...
I think ur solution was great.
First, Bogofilter will filter incoming messages.
if spam is detected, then it will be sent to the spambox ( i hope so),
then, i 'll try to check if this is real spam with SA. if so, SA will teach BF with this message.
what do u think about that ?
My problem is... how to write a shell script ( will be in the crontab ) to launch SA, scan the directory, and if results are good, send it to bogofilter...i'm quite bad at shell scripting :(
Thx a lot,
Config : BF + SA + Postfix.
Krispisen
On 3/22/06, mouss <[EMAIL PROTECTED]
> wrote:
Krispisen a écrit :
> i wonder then how can i use both to definitively not receive spam....
> coz, i'm an administrator ( lots of computers on my network), and
> spamassassin is too slow...
> bogofilter seems to be the godd solution, but i'd like itt to learn
> automatically what spam is...
> that why SA can teach him... but how... ?
>
depends on your mail platform.
what you can do is use SA and BF by default:
- first filter through BF
- if BF "confidence/probability" is high enough, don't filter through SA
This step may be implemented in the MTA (such as postfix header_checks)
or in an MDA.
- if a message is detected as spam by SA, then run it through BF's
learning command. This is the debatable part. Later on, if the message
was an SA False positive, you'll need to retrain BF (and SA if using
SA/Bayes|AWL). This step can be performed by an MDA (maildrop, procmail)
for instance.
but bayesian filters need "feedback" to correct their decisions. you
can have a list of "trusted" users that feed back missed spam and false
positives. or you can implement per user Bayes, but this will only be
good for those users that really feed the filter back. so it really
depends on your user community.
If you have a spamtrap (a safe one of course), use it to feed BF (and SA
if you enable SA/Bayes|AWL)
Initially, you can train BF (and SA/Bayes) with a corpus if you have
one. If you don't, you may try a public one (SA corpus, Enron corpus,
...). Google will get these...
