SpamAssassin - Bayesian Filtering

Bayesian Filtering

SpamAssassin by default tries to reinforce its own rules through Bayesian filtering, but Bayesian learning is most effective with actual user input. Typically, the user is expected to "feed" example spam mails and example "ham" (useful) mails to the filter, which can then learn the difference between the two. For this purpose, SpamAssassin provides the command-line tool sa-learn, which can be instructed to learn a single mail or an entire mailbox as either ham or spam.

Typically, the user will move unrecognized spam to a separate folder for a while, and then run sa-learn on the folder of non-spam and on the folder of spam separately. Alternatively, if the mail user agent supports it, sa-learn can be called for individual emails. Regardless of the method used to perform the learning, SpamAssassin's Bayesian test will subsequently assign a higher score to e-mails that are similar to previously received spam (or, more precisely, to those emails that are different from non-spam in ways similar to previously received spam e-mails).