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

On Thu, Oct 16, 2008 at 8:23 PM, Bill Yerazunis <[email protected]> wrote:
> Well, if you want to hack, sure.  It's not too bad for Markov/OSB
> because those operate with a one-token-at-a-time pipeline.

Well, one correction needed there: /only/ if you stick to <unigram>.
Otherwise you have that word-hash-set --> feature-hash mixing function
like VT offers, only inlined in the classifier code.

If you want the gory details, seek out the source code for those
antique classifiers and check where Bill used hctable[] in there. :-P

"It's VT, Jim, but not like you know it."


> So, if they ask you for what was significant in hyperspace, show them
> the high-scoring example texts, because _those_ truly are the
> significant elements (and NOT the individual words or phrases within those
> texts).

Fun!



So the moral of the story is, when asked 'why?' when you're applying
statistics or probability:

"Well, ma'm, sir, do you want your baby to be maybe a boy, maybe a
girl, orrrr.... do you wish for a perfect John Doe .8 second kid?
(John Doe has 1.8 kids, at least over here, and unlike the Dow Jones,
John is going down less and less. ;-) ... But that's a whole another
story.)
Yep, you get that kind of answers.

Or put another way: use probabilistics (and Bayesian is just that) and
you entered the grand hall of koosher/halam gambling, meant for all
those sinners that are convinced they can hide their sinning behind
complex looking math. You also don't ask your croupier 'why?' when the
ball doesn't end up on red 18 while you just thought 18 was your lucky
number this evening. All you got is hope and then there's that elusive
'trust' the house isn't skimming too much off the top. And since you
are trying to sell this to customers, you're the croupier. In
gambling, 'why?' is impolite to ask.

'Would you help us improve our betting?' is a much, much nicer question.

... And then grand master Bill would say: "Yep. Can do. Just train the
bugger the /right/ way and you're going to have a /great/ evening!"
(okay, there's still a few folks with headaches out there, but there's
always a few sore losers at the table.)


Pun intended, but you really should seriously consider educating your
customers into asking that much, much nicer question instead, before
they start using these statistical gynmastics. It's not for the 'wanna
be sure' guys and dolls if you get my meaning.



-- 
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=/