The SpamAssassin software includes
a "Bayesian" probability classifier, which is a database that
"learns" from SpamAssassins decisions, about what is spam and
what is legitimate email (often referred to as "ham").
This database of word patterns will then be used to improve the accuracy
of subsequent decisions.
If the spam score is very high, or very low, Spamassassin will
automatically run the message through sa-learn, and it will be
reported in the X-Spam-Status:
email-header as
autolearn=spam or autolearn=ham.
If spamassassin gets it wrong, you can correct it by updating
the status manually.
Sometimes, however, SpamAssassin doesn't have enough information to decide
which category a specific email belongs to, flagged as
autolearn=no in the X-Spam-Status: header,
and in those cases you might want to teach it what to do.
The procedure for feeding the bayesian learner is, to put a
copy of the message(s) you want it to learn from,
including all headers, into either
/usr/spool/mail/spam or
/usr/spool/mail/nonspam directory, using a filename equal
to your own username.
The next time you receive email (whether ham or spam), the learner will pick up the message(s) you copied there, and delete them after processing.
References: