> -----Original Message----- > From: Matt Kettler [mailto:[EMAIL PROTECTED] > Sent: March 2, 2006 8:53 AM > To: [EMAIL PROTECTED] > Cc: users@spamassassin.apache.org > Subject: Re: question on training spamassassin > > Webmaster wrote: > > >> Also if your users are only or mostly forwarding spam, > SA's bayes is > >> going to have a bayes bias that all messages forwarded by > your mail > >> clients are spam, regardless of content. > >> > >> > > > > Does this also mean that it is almost useless to share > bayes from one > > server to the next if each server has its own set of hosted > domains ? > > Yes, it's definitly very sub-optimal to share bayes DB's > across different domains, but not for the reason of header > differences. > > The reason this is useless is that the nonspam mail received > by different domains is not likely to be similar. > > Take for example a shipping company and a law firm. How much > similarity is there going to be in the day-to-day nonspam of > these sites? Sure both are likely to have some personal "Hi > hon, working late, be home at 7pm" type emails. However their > commercial nonspam is going to be VERY different. > > > > Because if the headers play such an important role, spams > targetting > > different sets of domains, I assume, are learned differently. > > To some degree, yes, but this is less severe than forwarding. > > At least things like source IP, User-agent, Message-ID and > other patterns are NOT going to be different across domains. > > To: and Received: headers will be considerably different, but > with a forward you retain ZERO of the original headers. >
ok gotcha! At least it will not be entirely useless. I am assuming that the non-spam, false positive issues will not be severe. Thanks.
