On Mon, 5 Jan 2026, Jerry Malcolm wrote: > I realize that the main function of SA is to separate spam from > non-spam. But I am encountering a growing need to sub-divide some > non-spam mail between important (keep) specific mail related to a > purchase vs. promotional mail from the same vendor. For example, we do > a lot of RVing. We make reservations sometimes months ahead at a > campsite and receive confirmation emails, etc. But we also like to hear > about other promotions and offers from the same camp site. I want to > expire and autodelete promotions-only email after a certain expiry > period while permanently keeping reservation confirmations/receipts. It > would be wonderful if these emails came from > "[email protected]" and "[email protected]". But > that would be too easy, and rarely occurs. Mostly everything comes from > info@... or some other common email. > > I suspect this is a total long-shot. But just curious if there are any > tricks/techniques that others are using that can somewhat definitively > separate reservation confirmations from promotions/marketing email so I > can categorize lifecycle accordingly? I suspect there may be some > training per vendor. But what is the best/recommended service to be > trained? Any suggestions welcome.
Maybe I am understanding Jerry differently than you other folks on the list. spamassassin comes with a set of rules that are tailored for finding spam. The trained bayes rules are just one of the rule types. In the end, spamassassin adds up a score of all the rules that trigger to let you make the spam/ham decision when that score passes a threshold. This isn't suited to classify mail into more than a 2 categories. You wouldn't just try to train spamassassin on your important and unimportang mailbox, this isn't going to work because all the other rules matter too. I suppose you could entirely gut spamassassin's rules and create your own rules and then train the bayes on your own stuff but you are kinda out on your own and frankly this would be a lot of work. I can't help but think there are easier ways to do this like some of the others on the list suggested. If you really wanted to do this by training something, I bet it wouldn't be too difficult to write some something say in python using something like tensorflow. Michael Grant
signature.asc
Description: PGP signature
