Disable Bayes-Filter?

[wwweiss]

My system often shows a bayes value of -1.9 on mails that are obviously spams. Sometimes this negative value is the reason why some spam mails are no spam.
I do not have any function to teach the bayes filter. Should I disable bayes filter?
What do you recommend to get the bayes filter working how it should?

 

[Stoiko Ivanov]

My system often shows a bayes value of -1.9 on mails that are obviously spams. Sometimes this negative value is the reason why some spam mails are no spam.
I do not have any function to teach the bayes filter. Should I disable bayes filter?
What do you recommend to get the bayes filter working how it should?

Hm - from my experience I would recommend to simply turn off the Bayes filter and see if you get acceptable results without it:
* From my experience (which is of course biased, since the support requests and questions in the forum seldomly are focused on things that are working fine) bayes with autolearning leads to more false classifications than it helps.
* autolearning (see https://cwiki.apache.org/confluence/display/SPAMASSASSIN/BayesFaq) is the only training implemented directly in PMG

In general I would suggest to implement the recommendations from our getting started page (which mostly correspond to what we run here and which works well for us):
https://pmg.proxmox.com/wiki/index.php/Getting_started_with_Proxmox_Mail_Gateway

Should you still want to manually train the bayes filter - you can use sa_learn from the commandline to do so.
should you do so - we would be very grateful if you shared your experiences (with some logs and numbers how it did perform)

I hope this helps!

 

[wwweiss]

I already tested with Bayes turned off and the results are definitely better. Some years ago I made serveral tests on another mailsystem with the bayes filter. I found that teaching the filter works good if you have enough spam AND no spam (at least 100 mails each before running the as_learn).
This is the basic problem that you easily get 100 spam mails from your customers but you do not get a good set of no spam.
You may offer some folders for the users where they can put spam and no spam but they only put spam.
Teaching the filter with only a lot of spam but only a handful of no spam results always gave negativ results (many spam mails got negativ points and thus went through).
I also had no good experince with the autolearn feature, which I can understand. Maybe we are getting good made spam mails offering some pills that are not detected as spam, then the autolearn of course will result in defining these mails as good.
Overall my experience also says: do not use Bayes if you are not able to teach the system with a good selection of spam and no spam. Do not rely on users filling your learning folders with mails.
Of course this is also just my personal experience which might be different in other situtations.