Re: Any way to tell what features contributed the most for a particular decision?

Short answer: no

Longer answer: for hashtable-storage-based classifiers you can dump
(cssutil backup IIRC) the CSS files (OSB, OSBF, Winnow, Markovian) and
thus get one part of what you need: the learncounts for individual
features.

Multiple hurdles to take from there: features are hashes. crm114
doesn't keep around the words that turned into those hashes as there's
not use for 'remembering' those. (unless you're in deep diagnostic
modes which are way out of the ordinary).

latest experimental GerH builds have a profiling mode which stores the
original text (partly), but that only half the story there: words are
mixed to produce feature hashes so that complicates matters even
further. plus the experimental code is quite unfinished.

lastly, there's the classification math implementation turning feature
hits into scores: there's nothing in there which would easily lead to
a 'TOP x' ranking of individual features, unless you went at crm114
with a BFG9000.

So, all in all, no. Unless you have a math geek available and
development time for /major/ crm114 hacking.


The hackish alternative I can come up with is using multiple CSS files
for both sides (not just two, but many) and only learn certain words
into certain CSS files, so that scores for CSS would directly
represent the hitrate for certain known (sets of) words for the given
samples.
Assuming a small size word set, it might be doable. Maybe use multiple
rounds of CSS construction, each time put a different mix (i.e. remix)
of words in a CSS, then reclassifying all samples, so you can deduce
after combining all results, which 'words' have the highest
probabilities of significantly contributing.
Might work in theory, relatively small word sets only, but it's ugly
and my math betters will surely be able to point out fatal flaws in
the basic idea.





On Thu, Oct 16, 2008 at 1:10 AM, Robert Swirsky-Warner
<[email protected]> wrote:
>
> When using crm114 to 'classify', is there any way to get a report on what
> the top features were that contributed to a particular classification? I'd
> like to know what words or phrases were the most significant to the
> classifier for any particular test.
>
> --
> Robert Swirsky
> Theme Park Photo, LLC
> Sunnyvale, CA
>
> -------------------------------------------------------------------------
> This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
> Build the coolest Linux based applications with Moblin SDK & win great
> prizes
> Grand prize is a trip for two to an Open Source event anywhere in the world
> http://moblin-contest.org/redirect.php?banner_id=100&url=/
> _______________________________________________
> Crm114-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/crm114-general
>
>



--
Met vriendelijke groeten / Best regards,

Ger Hobbelt

--------------------------------------------------
web:    http://www.hobbelt.com/
       http://www.hebbut.net/
mail:   [email protected]
mobile: +31-6-11 120 978
--------------------------------------------------

-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/