Bayesian mail filters -- such as POPFile -- categorize mail according to what they have learned in the past. If they first see spam, they don't know it, but if you tell a Bayesian filter that a particular message is junk, it will learn from that experience, and if it determines another mail to be spam based on this experience it will learn some more.
That's why it is essential that POPFile learns from both good and bad mail. It knows each bucket only by the examples it has seen so far. If you use a magnet to bind mail from somebody to a certain bucket, POPFile does not learn from that person's messages. If somebody else (or the same sender using another address) sends the same message, POPFile may not know where it belongs or make assumptions that are completely wrong.
Of course, using magnets can be useful, but the scenarios -- the friend that keeps forwarding you spam and you like it -- are scarce. Usually, it's best to stay away from magnets in POPFile and just teach it as it makes mistakes (which, at about 99% accuracy isn't that often).
