Am 08.01.2015 um 22:57 schrieb Alex Regan:
On 01/07/2015 02:31 PM, Reindl Harald wrote:Am 07.01.2015 um 20:23 schrieb Alex:I'm also wondering what exactly you're taking from these messages that are received? Are you blocking based on IP? Creating header/body rules? Those are usually transferable to other systems, but what about bayes? How can you use it for bayes when that doesn't transfer very easily to other systems?depends how you build your bayse mine is a phyiscal ham and a spam-folder with raw-mails and a script calling sa-learn which makes it easy to transfer to other systems as well as maintain the corpus over yearsIt's not necessarily an issue with the bayes database itself, but rather the content of the database - emails comprising your ham database may not constitute ham for my company, for example.
your opinion mine is * if it is ham and handselected than it is ham * if it is spam and handselected than it is spamhence the *global* bayse i maintain for a few months for 1500 users with very different businesses with the help of some of them works that much better than the previous "light maintained" global and the mostly completly useless user-bayes which don't reach the 200/200 balance or are trained by people not understand how a bayes works
at the end of the day you need *as much as possible* samples for ham and spam where you trust the classification - currently newsletters are not marked as spam while a lot of similar *looking* messages are trustable blocked at milter level
* no databases * no magic * no auto learning * just a global bayes with a corpus of 15000 hand-selected mails in the last 5 months with a nearly perfect hit-rate and zero false-positives highly scored
signature.asc
Description: OpenPGP digital signature
