On 15-Mar-2009, at 02:29, decoder wrote:
I'm thinking that FPs and FNs are bayes problem anyway. This tool need to concentrate on seeing just what rules hit and building off that. I'd go so far to say that as far as SVM is concerned, there is no such thing as a false postive or negative.
What do you mean by that? Of course FPs and FNs might also be a problem for the SVM, every wrong classified point is certainly a problem for a machine learning algorithm. However, I think that the SVM is quite robust to a certain amount of FPs/FNs if the majority of the training points is correct.
Well, if the data you feed it has a lot of False positives/negatives than of course your data us going to be skewed. But that is still the responsibility of bayes/SA which generated the false positives. If you have a spam tagged mail that is not spam and you feed it to your HAM position of SVM, then SVM doesn't care that it was a miss-tagged, it's going to weigh that with the other ham messages. Similarly with 'ham' message that is actually spam and gets fed in as spam.
So, if you feel like trying out the plugin, let me know how well it works =) I'm especially interested in those cases where it increases the spam detection rate (reducing false negatives). Might be easy to extract this information from logs.
I'd love to try it out, but I am in the midst of making a lot of changes to my postfix config as it is, so it's going to be at least a week or two before I can think about implementing it.
-- Hamburgers. The cornerstone of any nutritious breakfast.
