Adaptive weighting? [LWN.net]

Adaptive weighting?

Posted Mar 4, 2004 3:02 UTC (Thu) by Ross (subscriber, #4065)
In reply to: Keeping spamassassin current by bronson
Parent article: Keeping spamassassin current

Why can't they do some kind of Baysian network based not on single words
from the text but on the output from the various SpamAssassin rules? Then
they wouldn't have to fine-tune the weights and individual users would get
weights that matched the spam in their inboxes better. This would also
make it more difficult for spammers to test against the rule base since
they wouldn't know which rules are weighted heavily and which lightly.

(Log in to post comments)

Adaptive weighting?

Posted Mar 4, 2004 4:34 UTC (Thu) by proski (subscriber, #104) [Link]

Sounds like an excellent idea! I hope you will share it with Spamassassin developers. Complex rules is Spamassassin's strength. The way how they are combined (addition) is spamassassin's weakness. Predictiveness of the score is another weakness. Let's get rid of weaknesses.

Adaptive weighting?

Posted Mar 4, 2004 6:09 UTC (Thu) by mkettler (guest, #3933) [Link]

SpamAssassin has had a bayesian filter, in addition to the rules, for the past 8 releases. The first version with a bayes subsytem was 2.50, released Feburary of 2003.

No need to share this with the sa-devs.. they clued in a long time ago.

Adaptive weighting?

Posted Mar 4, 2004 20:23 UTC (Thu) by skybrian (subscriber, #365) [Link]

It sounds like you misunderstood the point. Unless something changed since 2.50, the bayesian filter is just a separate set of rules. The weight on each rule (including the Bayesian rules) is static.

Adaptive weighting?

Posted Mar 4, 2004 10:33 UTC (Thu) by nix (subscriber, #2304) [Link]

This has been tried, and wasn't terribly effective.

(Justin posted some test results on this to the sa-dev list maybe a year ago.)