Robinson-Fisher (aka chi) beats Bayes chain rule
Greg Louis
glouis at dynamicro.on.ca
Sat Nov 30 03:49:40 CET 2002
A correction (sorry) and some further information:
On 20021129 (Fri) at 2039:57 -0500, Greg Louis wrote:
> A detailed writeup of this experiment will appear sometime during the
> coming weekend at http://www.bgl.nu/~glouis/bogofilter/FisherBcr.html
> but for now here's a summary:
>
<snip>
>
> We can calculate the means and 95% confidence limits of these values,
> and they come out like this:
>
> calc meanfnpc lcl95 ucl95
> 1 chi 1.51 1.45 1.56
> 2 bcr-fw 1.79 1.73 1.85
No they don't; I entered the wrong number of degrees of freedom for the
residuals (4 instead of 2). The actual confidence limits are these:
calc meanfnpc lcl95 ucl95
1 chi 1.51 1.38 1.63
2 bcr-fw 1.79 1.66 1.92
A corrected graph, file bcr2.png, is attached.
Also, it turns out that I can force the false-positive count down to 71
and still get a meaningful comparison (remember, chi is capable of
operating with a cutoff high enough that only 42 false positives are
reported). The result of doing this is interesting: chi's
false-negative count climbs only a little, but that of the chain-rule
method is much more strongly affected:
calc run fneg percent
1 chi 0 30 2.118644
2 chi 1 25 1.765537
3 chi 2 27 1.906780
4 bcr-fw 0 91 6.426554
5 bcr-fw 1 84 5.932203
6 bcr-fw 2 96 6.779661
calc meanfnpc lcl95 ucl95
1 chi 1.93 1.25 2.61
2 bcr-fw 6.38 5.70 7.06
The second graph, bcr2a.png, shows this result.
--
| G r e g L o u i s | gpg public key: |
| http://www.bgl.nu/~glouis | finger greg at bgl.nu |
-------------- next part --------------
A non-text attachment was scrubbed...
Name: bcr2.png
Type: image/png
Size: 2727 bytes
Desc: not available
URL: <http://www.bogofilter.org/pipermail/bogofilter/attachments/20021129/5ee5c93a/attachment.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: bcr2a.png
Type: image/png
Size: 2770 bytes
Desc: not available
URL: <http://www.bogofilter.org/pipermail/bogofilter/attachments/20021129/5ee5c93a/attachment-0001.png>
More information about the Bogofilter
mailing list