[prev in list] [next in list] [prev in thread] [next in thread] 

List:       spamassassin-users
Subject:    Re: R: Bayesian filter error?
From:       Axb <axb.lists () gmail ! com>
Date:       2014-10-29 10:47:18
Message-ID: 5450C5B6.5020600 () gmail ! com
[Download RAW message or body]

On 10/29/2014 11:38 AM, Marco Tironi / 8volante Srl wrote:
> Thanks for your fast reply. Now I understand the big mistake: Bayesian
> filter is server specific and not "public" so it's not globally manteined.
> Every server have its own indexes so there is no fast solution to solve it
> globally.
>
> I can allow that signrature for my server, but others server continue to
> mark them as spam.

You are using a dangerously low "required" threshold:

X-Spam-Status: Yes, hits=2.8 required=2.0

Apart form maybe having a mistrained Bayes DB, running such a low 
required level is asking for trouble.

The "fast" solution is to manually retrain your bayes with those FP'd 
msgs as ham and the rethink your rquired score.

Your problem is mainly local - what other servers do is out of your control.

> Marco
>
> -----Messaggio originale-----
> Da: Reindl Harald [mailto:h.reindl@thelounge.net]
> Inviato: mercoledì 29 ottobre 2014 11:25
> A: users@spamassassin.apache.org
> Oggetto: Re: Bayesian filter error?
>
>
> Am 29.10.2014 um 10:50 schrieb Marco Tironi / 8volante Srl:
>> Hi, I use a Spamassasin version 3.3.1 on Windows System and I have a
>> problem with Bayesian filter:
>>
>> -A legitimate users send an email to our server and they are delivered
>> normally
>>
>> -When that users insert it's domain in the email signature the email
>> is marked as spam with that header:
>>
>> oX-Spam-Status: Yes, hits=2.8 required=2.0
>> tests=BAYES_99,BAYES_999,HTML_MESSAGE,RCVD_IN_DNSWL_LOW,SPF_PASS
>> autolearn=no version=3.3.1
>>
>> oX-Spam-Score: 2.8
>>
>> -After a few try I have noticed that if I remove the domain line form
>> the signature the email is delivery correctly
>>
>> -That users report that the problem is the same on many server (I
>> think that's because all use BAYES filter
>>
>> How can I inform the manteiners of Bayes Filter of this false positive?
>
> normally i would expect that behavior triggered by URI blacklists which did
> not hit in the tests - if it is really bayse than it means a lot of messages
> containing this domain are trained as spam and none as ham
>
> a bayse relies on proper training
>
> meaning you need at least the same amount ham samples for a really good
> working one as you have spam samples
>
> what says "sa-learn --dump magic"
>
>


[prev in list] [next in list] [prev in thread] [next in thread]