Hi, On Sun, 15 Jan 2006 14:07:51 +0000 Stroller <[EMAIL PROTECTED]> wrote:
> On 15 Jan 2006, at 12:56, Hans-Werner Hilse wrote: > > > > I'd strongly suggest using the Bayesian filters, per-user, that is... > > [...] > > What improvement rate are you seeing for this, please? About 99% of _Spam_ mails are positively recognized. Up to now I've never encountered a false positive. > My concern with these particular users, who are not particularly > email-savvy, is that they ain't going to train the filters. I just > don't see it happening. And if I teach them to train the filters by > dragging & dropping into the "learn" folder then I anticipate perhaps > just one of them complaining "but why can't I just right-click it and > `mark as junk' in Outlook?". True. My answer is: "because then you'll get all the spam in your webmail when being on business trip" :-) Basically, I teach them to use server-side mail filtering with the same reasoning. But it makes me think: Does Outlook set some kind of flag to the mail? Does it note anything in the headers? > I'd really prefer all spam-filtering to be invisible to the user. I > don't demand a high success rate: Bayesian filtering should get 99.5% > or above, I think, but I'd be happy with 95%. In fact, lots of my users are happy with about that rate and without learning of Spam. Bayesian filters are activated for all of them, but they are only trained by autolearning. > SpamAssassin is currently getting about 33%, which is next to useless. agreed, and I bet you can improve that. You can also decide to have all users share your Bayesian database. So you don't have to teach them to learn Spam. > IMAP server is Dovecot storing messages in maildirs in users' home > directories - this makes it convenient for your suggestion, but I > just don't really want to go there. You can, as described, reduce the concept at many points... -hwh -- gentoo-user@gentoo.org mailing list
