bogotune results

[Peter Bishop]

On 24 Mar 2004 at 21:43, David Relson wrote:

> They're interrelated in non-obvious ways.  Simply increasing one (or
> all) of them doesn't lead to a clear increase (or decrease) in scores.
> If they were related in some obvious way, we'd have a set of equations
> to spit out the ideal parameters.  As it is, we have a program that does
> some heavy duty number crunching and searching in order to come up with
> a good set of parameters.

Perhaps there is some scope for testing for relationships
for individual messages rather than "in bulk".

Here is something I did in about June last year. 
(see attached graph) using an old version of bogofilter.

The y axis is spamicity is measured using the Robinson GM measure
which is more linear than Fisher 
- so you can see impact of changing the parameter more easily.

The spam cutoff point is 0.54

The x axis values are just arbitrary spam reference numbers

And the graph shows how the spamicity of these emails varies for 
different values of the mindev parameter.

As you can see, there is a fairly consistent increase in spamicity 
as mindev increases for all the individual emails. 

Perhaps a more comprehensive study could be performed along these 
lines?
-- 
Peter Bishop 
pgb at adelard.com
pgb at csr.city.ac.uk


-------------- next part --------------
A non-text attachment was scrubbed...
Name: mindev.gif
Type: image/gif
Size: 4967 bytes
Desc: not available
URL: <http://www.bogofilter.org/pipermail/bogofilter/attachments/20040325/ea632e48/attachment.gif>