As spam is constantly changing, the Bayesian filter must be kept up to date. Keep the following in mind when using the auto-learn feature of SpamAssassin:
The auto-learn feature should be configured to only learn from the emails that are definitely ham or definitely spam, but not the ones in between the two auto-learn thresholds. Emails with very low or very high scores are the ones that SpamAssassin can detect without Bayesian filtering. It is the middle ground that Bayesian filtering can assist with most. By omitting these emails from the Bayesian learning process, the effectiveness of the Bayesian filter is decreased.
There is a chance that emails may be mistakenly tagged if auto-learn alone is relied upon. If a spam is taught to SpamAssassin as a ham, or vice-versa, then this will adversely affect all future Bayesian results.
For these reasons, it is recommended that a policy of regular manual training is used.
It is important to teach SpamAssassin a varied selection of...