Quanah Gibson-Mount wrote:
--On Wednesday, April 15, 2015 3:34 PM +0200 Mark Martinec wrote:
Don't know. It might be worth taking a look at the CRM114 classifier,
which implements a number of methods. The CRM114 can be used as a
plugin to SpamAssassin, or can be called from Amavis and results
combined with those from SpamAssassin. We had good results from
CRM114,
and some friends of mine are very enthusiastic about it. Certainly
it is a good complement to SpamAssassin's naive bayes classifier,
I was looking at adding CRM-114 to Zimbra to integrate with SA and/or
Amavis as a dspam, but it seems to have the same issue dspam has --
Abandoned. The last mail to the announcement list was in 2007 about
some issue with the software, and prior to that was an email in July
2006 about a new RC. It's been dead silent since. There's some minor
traffic on the users list.
It also depends on a program called tre, which also appears to be
abandoned, although it at least had a commit in the last yearish.
That is unfortunately true, it seems to be abandoned. It also uses
some unorthodox implementation language. But it does work as advertised.
I haven't noticed any malfunctions, and its results are comparable
to SA's Bayes: when they agree it adds valuable score points to marginal
spam, and when they disagree it is not unusual that one or the other
saves us from a false positive.
CRM114 can be a rich source of ideas and AI algorithms, which is why
I mentioned it in this thread - worth looking into its approaches.
I would be interested to hear more about what results your friends
are seeing and how they are integrating it into their workflow.
Don't have any statistics on that. It's been in use as either the
SpamAssassin plugin or as Amavis external spam filter - either way
it does its job, and one or the other approach has some benefits and
drawbacks. I believe auto-learning-on-error is in use (after initial
training) through learn_ham and learn_spam fields in an @spam_scanners
entry in amavisd.conf.
I was looking at adding CRM-114 to Zimbra to integrate with SA and/or
Amavis as a dspam, but it seems to have the same issue dspam has
It may not be the best match for an unattended operation at some
remote customer location, I agree. It uses a fixed size database
which may need occasional resizing. For a managed site with a
knowledgeable administrator it can be a good fit - it certainly
does well for some sites. But as with all auto-learning approaches:
the quality of results depends on the quality of input. Better classical
rules (DNS, regexps, ...) yield better auto-learning, and the more
homogeneous a use base is (with their mail content), the better
results are.
Mark
