Hi, I actually think it is more to do with the fact that one person's spam could be another person's ham. If the mail streams and servers are carrying messages for a community of users who receive (and want to receive) similar types of email messages, I can't see any major problem with using those emails to train Bayes. However, if the servers are processing email for two completely different user communities their ideas of what is and isn't spam could be so different that the Bayes stats become diluted.
For instance I work for a Medical School, but in a heavily IT based department. Some terms that may be considered pornographic for someone working in banking could be perfectly acceptable in my environment. > If I am understanding this correctly...the concern is that the Bayes > should match the mail server in which the ham and spam was received on > only? > > David Roth > rothmail (at) comcast.net (dot) net > > On Nov 18, 2005, at 5:10 AM, qMax wrote: > > > in wiki://BayesInSpamAssassin it is said: > > Do not train Bayes on different mail streams or public spam corpora. > > These method will mislead Bayes into believing certain tokens are > > spammy or hammy when they are not. > > > > Could you explain why it is so, and what could happen if to teach > > nayes from several mail servers ? > > > > -- > > qMax > > > -- Anthony Peacock CHIME, Royal Free & University College Medical School WWW: http://www.chime.ucl.ac.uk/~rmhiajp/ "Computer software consists of only two components: ones and zeros, in roughly equal proportions. All that is required is to sort them into the correct order."
