Re: spam-stat.el 0.0.4

[Alex Schroeder <alex@emacswiki.org>]

Alex Schroeder <alex@emacswiki.org> writes:

> ;; I used 323 mails in my spam directory, and 180 mails in my mail
> ;; directory.  SpamAssassin is installed on my server.  The result of
> ;; my tests:  Testing spam -- 293 / 324 = 0.904321 (10% of the spam
> ;; missed), testing non-spam -- 6 / 180 = 0.033333 (3% false
> ;; positives!).

Ich checked these 6 false positives (this would be a major problem):

1. only one mpg file
2. bounced mail, includes three jpg files
3. only one mp3 file
4. this really is spam
5. this is really not spam at all -- a portuguese mail
6. only one tar file

No 5 can be explained by the very small number of portuguese mails in
the sample.  Here is the debug output.  As you can see, the common
words "para" and "esta" account for most of the high scores.


spam-stat-score-data's value is shown below.

Documentation:
Raw data used in the last run of `spam-stat-score-buffer'.

Defined in `/home/alex/elisp/spam-stat.el'.

Value:
(("para" 0.99 0.49)
 ("para" 0.99 0.49)
 ("esta" 0.99 0.49)
 ("dia" 0.01 0.49)
 ("para" 0.99 0.49)
 ("Sascha" 0.01 0.49)
 ("schroeder" 0.01 0.49)
 ("Sascha" 0.01 0.49)
 ("esta" 0.99 0.49)
 ("para" 0.99 0.49)
 ("Hm" 0.01 0.49)
 ("que" 0.99 0.49)
 ("que" 0.99 0.49)
 ("Porto" 0.01 0.49)
 ("para" 0.99 0.49))


ATM I do not feel too bad about the false positives.  I hope that
these will disappear with a larger training set.