[Spambayes] "Un-classifying" a message

[Tim Peters]

[Tim Peters]
>> Speak for yourself <wink>.  I keep all my training ham and spam in a
>> distinct .pst file dedicated to holding training data, and I retrain
>> from that.  There's no mistake I can't recover from easily that way.
>> I've also

[Seth Goodman]
> OK, but if I understand this right (50% odds at best), you have to
> manually copy each message you train on into these folders.

That's 50% right <wink>.  The addin sends spam directly to my trained Spam
folder, leaving it marked unread.  At the end of the day (or every two days,
...), I *delete* the unread msgs in the Spam folder (after reviewing for
mistakes).

I do copy ham to the ham training folder.

> If you are talking about the original training set, I agree with you
> and I do the same for exactly the reasons you gave.  If you are talking
> about training during normal operation, that's a lot of work with a
> lot of room for manual errors.

I expect it's different for everyone.  My training database right now
contains about 800 emails total, a little more than one day's worth.  I
don't find much need to train, and the manual error rate in dragging a
multi-selection of Unsure ham to the training folder hasn't been high enough
to care about.  About once a week I rescore my training folders and look at
"the wrong end" of each for manual mistakes; they're rare; hasn't happened
for months.

Because everyone's email mix is different, and tolerance for fiddling with
training varies widely too, and ours is a statistical approach, I don't
expect that a one-size-fits-all strategy can exist.

...

>> A simpler solution (one I wouldn't use, but ...) would probably be
>> to add a single new checkbox option analogous to the existing "train
>> on move" options:  untrain a message if it's moved to the Unsure
>> folder from a Ham or Spam folder.

> Yes, this would be simplest, but suffers from the same "invisible
> action" on folder move that we both don't like.  How about a button
> that says, "Recover to Unsure"?  At least it's unambiguous.

Well, the toolbar is already so wide that we get a lot of bogus bug reports
about one of the (just two!) buttons on it "missing" (Outlook simply doesn't
show it because there's not enough screen real estate).  Needing to untrain
a message is (or should be) so rare that I'd settle for an "Untrain
selected" action on the dropdown SpamBayes menu instead.  Buttons should
really be confined to frequent actions.  A really gonzo approach would
follow Outlook's lead, creating an icon for each possible action and letting
users populate their SpamBayes toolbar with whicever buttons they want.
Without much effort along those lines, we could make the SpamBayes codebase
bigger than Outlook's <wink>.