the mozilla implement is based on Paul Graham's A Plan for Spam.
training data is global across all mail accounts within a profile
training data is stored in a binary format, in a file named "training.dat".
initially, the training.dat file is empty (there was discussion of shipping with a default file)
on spam detection, the user can choose to move spam to a special "Junk" folder
the user can configure junk mail can be automatically purged from the "Junk" folder
to analyze a message for spam, we need the entire message, not just the headers.
spam detection happens after filters are run
white listing happens after filters, but before spam detection
the purge code is implemented as a search of the "Junk" folder, looking for "old" message that have the proper junk status.
when does purging happens?
elaborate on the mime changes that were made for spam
currently, spam filtering is does not work for news, but it would be possible to add support for this. (there is a bug on it.)
| initial
state |
user
action |
table changes |
| unknown (user can't see this, looks like "not junk") | mark as junk |
add tokens to bad |
| unknown (user can't see this,
looks like "not junk") |
mark as not junk |
add tokens to good |
| not junk |
mark as junk | remove tokens from good, add
tokens to bad |
| not junk |
mark as not junk | no op |
| junk |
mark as junk | no op |
| junk |
mark as not junk | remove tokens from bad, add
tokens to good |