[Spambayes] training suggestions

[Dhaval Patel]

> My fault.  Yes, if you want to do incremental training with sb_mboxtrain,
> then leave off the -f flag.  The fact that you are training once a day threw
> me though.  If incremental training is what you want to do, why not run
> sb_mboxtrain more frequently than daily?

I just trained manually and realize that it says trained 190 out of 190 messages even
though there are only ~20 new messages since the last training. Is the output wrong?

My problem with running it at all is misclassification. How do I get it to fix the ones
that are misclassified?

> 
>     Dhaval> Now that I think about it, the best thing would be to do the
>     Dhaval> following:
> 
>     Dhaval> 1. message comes in to the filter
>     Dhaval> 2. filter sorts it as ham or spam
>     Dhaval> 3. db is trained with this message ham or spam (just this one
>     Dhaval>    message) when the user sorts messages (put spam from inbox ->
>     Dhaval>    spam folder)
>     Dhaval> 4. the changes to the db made in step 3 get undone and the
>     Dhaval>    message gets trained as spam (similarly if the user moved
>     Dhaval>    form spam folder -> inbox)
> 
>     Dhaval> Can spambayes do this? Can I specify just a message id or a list
>     Dhaval> of message ids? ( I use maildir format mail storage)
> 
> I'm not sure.  Again, that's not the way I work.  Nothing is ever trained
> automatically in my personal environment.  If I see a message that is
> misclassified (either false negative, unsure or false positive) then I toss
> it into the appropriate training database.  I never train on a message which
> was properly classified. 

According to your last sentence, when I leave out the -f, I would also not train the
message that was properly classified right? Or maybe not. How would/should incermental
training be handled if the previous training occured with misclassifications?

> 
> Note also that I have the luxury of having a user population of one person.
> Sounds like you aren't so fortunate.

My users are pretty good though. They dont have annoying questions or problems. :)


Thanks,
Dhaval



--