I agree with your analysis of the problem. Why it's happening, I can't say - I'm sure Thomas will chime in, but in the interim, if you're not fining HMM reliable (please check more than just the 1 message) consider turning it to monitor mode instead of scoring/blocking so that the HMM inaccuracies don't cause poor results, at least that's what I would do.
I went through this when first turning on HMM. Turns out that the corpus wasn't great, an bayesian wasn't perfect either. I use subject name logging which helps my identify incorrectly classified messages in the corpus. I sorted the spam, not-spam, and errors folders by name. Then I eyeballed the filenames, looking for obvious errors. I also used messages from ok mail an discarded that I manually reviewed and moved to spam/not-spam to help learn. It's a tedious process, but necessary IMO when the corpus is too far out of whack. I also run a block report for all addresses sent to me nightly, and review every single block. Of course this doesn't show bad mail that is delivered - I rely on users to report those, and they're starting to. Big help. Hope this helps you get the server running as it should. On Fri, May 29, 2015 at 9:27 AM, Miles Gaynor <m...@castlehoward.co.uk> wrote: > Folks, > I've noticed a large increase in spam recently, mainly from marketing > types sending me genuine marketing that I don't actually want. I assumed > that was because the messages were convincing but I sent one to > asspanalyze@assp.local<mailto:asspanalyze@assp.local> and this is the > final part of the message. > > > Bayesian Spam Probability: > combined probability: > > 1.00000000 - got 111 - used 60 most significant results > > > ________________________________ > > Hidden-Markov-Model Spam Probability: > combined HMM spam probability: > > 0.0000 - got 44 - used 44 most significant results > > > > If I'm reading that correctly, the Bayesian filter is working fine but the > HMM isn't agreeing with it. > Could you give me a clue? (yes, I'm pretty clueless). > Thank you, > Miles > > > > ________________________________ > Castle Howard Estate is a limited company registered in England and Wales. > Registered Number: 480214. Registered Office: Estate Office, Castle Howard, > York, YO60 7DA. This message is private and confidential. If you have > received this message in error, please notify us and remove it from your > system. > > ------------------------------------------------------------------------------ > _______________________________________________ > Assp-user mailing list > Assp-user@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/assp-user > ------------------------------------------------------------------------------ _______________________________________________ Assp-user mailing list Assp-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/assp-user
