Nathanael, > Assuming no corpus-feeding, if dspam is trained > soley based on SA scoring decisions, ... > ...it would seem to me that in the end you have trained dspam > (via SA) to make the same decisions that SA would have made. > This means two things to me: > 1) dspam provides no additional benefit when used with SA because it > will have been trained by SA to make the same decision that SA itself > would have made. > 2) having dspam decisions adjust SA scores just magnifies the SA score > in either direction, i.e. increasing the risk of FP's and/or raising an > already-spam score or dropping an already-clean score.
Your conclusions is correct. The auto-learning of dspam from SA is supposed to be just an additional feedback. To get good results with dspam one should train it manually as a primary mechanism. Still, for the lazy the current scheme works reasonably well, but it has the shortcomings you noted. Also the resulting positive feedback loop is a mistake, it would be better to change the current code to add the dspam score to the SA score by amavisd (instead of by a SA rule), which would prevent the positive feedback - the same way as is currently done within SA to train its Bayes db, which does not take into account the score from Bayes in training decisions. I'll put it on a TODO list, but if someone comes up with a patch, so much the better (as I no longer use dspam and won't be able to test it). Mark ------------------------------------------------------- This SF.Net email is sponsored by: Power Architecture Resource Center: Free content, downloads, discussions, and more. http://solutions.newsforge.com/ibmarch.tmpl _______________________________________________ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
