[Spambayes] PocketPC

[Mark Hammond]

> Mark,
>
> Of course, the step where I reduce to the 5 most
> misclassified before training has to be done on the desktop
> -- the Spam field isn't available in Pocket Inbox.

Actually, I wasn't going to make it that clever.

> Also, my initial test indicated that folder moves (such as
> dragging from Inbox to Junk or vice versa) from Pocket Inbox
> aren't seen by SpamBayes -- ActiveSync appears to move the
> messages outside of Outlook, without notifying the plugin.  I
> may have missed something here.

That is very likely.  Note that what I proposed was very simple - 2 special
folders that we watch for new messages.  When a message arrives in one, we
train and move.

This means that the whole '5 most misclassified' thing is skipped -
SpamBayes itself may grow smarter incremental training schemes, and this
would hook off that.  Another way to look at it: if '5 most misclassified'
truly was effective in this scenario, presumably it would also be effective
when not using the pocket PC.

To handle the second issue, we would have to watch these folders in the
background.  This means we wouldn't see the removal of the message from the
'real' SpamBayes folders, but only the arrival of the message in the new
folder - but we can handle that (and is indeed how we handle existing manual
classifications)

> Anybody:  If you do "recursive" training with only a few
> messages added each time, how long should it take to get
> reasonable results?  After three cycles with an increment of
> 5, last night's results were 2 misclassified ham and 531
> misclassified/unsure spam to only 30 correctly classified
> spam (I didn't get a count of the good ham).

This is a common question, but one we really can't answer without actual
testing over multiple data sets.

Mark.