On Thu, 11 Sep 2014 16:11:33 +0200 Axb wrote: > On 09/10/2014 11:19 PM, RW wrote: > > On Wed, 10 Sep 2014 20:57:35 +0200 > > Axb wrote: > > > > > >>> > >In practice this means that, without custom rules, ham can > >>> > >only be autolearned if it hits a DNS whitelist rule or > >>> > >RP_MATCHES_RCVD. > >>> > > > >> > > >> >from what I'm seeing is that it takes lower scored ham to > >> >autolearn ham. I don't use DNS whitelists and RP_MATCHES_RCVD is > >> >disabled > > To reach -1.0 it has to hit some negative scoring rules. If you > > look at all the rules with negative scores and eliminate rules > > marked learn, noautolearn or userconf you have the rules listed > > below. > > The point is to make auto learning ham less sensitive and it works.
There's a qualitative difference between a threshold of 0.1 and -1.0. At 0.1 ham can be learned just by not hitting any spam tests, at -1.0 it has to be picked-out by "nice" tests - this means that training could be too selective to be useful. This is something that can look great as a box-ticking exercise without actually being an improvement. If you don't understand how it working it's hard to tell whether it's working well, and much harder to address any shortcomings.
