Bayesian filtering

What it means

Paul Graham's 2002 essay A Plan for Spam popularised the approach: instead of hand-written rules, the filter counts how often each word or token appears in known spam and in known good mail (ham), and combines those probabilities to judge a new message. Because it learns from each user's or organisation's own mail, it adapts to what that audience receives.

Apache SpamAssassin ships a Bayes plugin described as a Bayesian-style probabilistic classifier based on Graham's algorithm. It is trained with sa-learn, and by default it does not activate until 200 spam and 200 ham messages have been learned, to stay accurate.

For senders this means content patterns that recipients have repeatedly reported can count against later messages that share them, even from another sender.