[prev in list] [next in list] [prev in thread] [next in thread] 

List:       wekalist
Subject:    Re: [Wekalist] Choosing a classifier for attribute selection
From:       "Su, Jiang" <k4km1 () unb ! ca>
Date:       2005-12-29 16:19:09
Message-ID: 1135873149.43b40c7da124d () webmail ! unb ! ca
[Download RAW message or body]

Hi:

I would like to recommend Bayes networks classifier(TAN is the best one in most
situation), and the second one is C4.5(If your datasets has numerical attirbute,
decision tree may be better). I am afraid that you have to wrtie some code to get the
score for each attribute in classifiers. You may use cross validation to relax
overfitting problem.

Happy new year.

Jiang Su

> 
> 	Hello all,
> 	I'm trying to identify (and even rank if possible)
> the attributes  
> and/or subset of (numerical) attributes that best
> predict a (nominal)  
> class. After reading the paper from Hall and Holmes
> ('Benchmarking  
> attribute selection techniques for discrete class data
> mining') I  
> decided to use a classifier-driven approach (part of
> the attribute  
> subset evaluation techniques) to do this, instead of
> something like a 'simple'
> gain ratio evaluation.
> 
> 	My question is : how to choose a classifier for this
> wrapper (ClassifierSubsetEval) ? The  
> obvious approach is to choose the one that 'best'
> predict the class  
> when taking all the attributes (I even developed a
> software embedding  
> Weka to evaluate all the classifiers able to deal with
> a given  
> dataset), but is that a so good idea ? Is there a
> potential over-fit  
> problem with this approach ? And if yes, then how to
> choose a  
> classifier ? On which criterion ? For now I use the
> fraction of  
> correctly classified instances as my 'score'. But this
> maybe  
> introduce a bias somewhere ?
> 
> 	If a machine learning guru is on this mailing list
> and have time to  
> consider this question, I will be glad to follow its
> advice. Thanks !
> 
> 	Aurelien Mazurie
> 
> (ps : of course if somebody think that a
> classifier-driven attribute  
> selection is silly and have an other suggestion, I'm
> open to  
> suggestions ;-)
> 
> (other ps : I wonder if a model exist to ... choose a
> classifier best  
> fitted for a given dataset. A kind of
> meta-metalearning ?)
> 
> ___
>      Aurelien Mazurie, Ph.D.
>      Post-doctoral fellow
>      Center for the Study of Biological Complexity
>      Virginia Commonwealth University
> 
>      http://oenone.net/contact/
> 
> 
> 
> 	
> 
> 	
> 		
> ___________________________________________________________________________ 
> Nouveau : téléphonez moins cher avec Yahoo! Messenger ! Découvez les tarifs
> exceptionnels pour appeler la France et l'international.
> Téléchargez sur http://fr.messenger.yahoo.com
> 
> _______________________________________________
> Wekalist mailing list
> Wekalist@list.scms.waikato.ac.nz
> https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
> 



_______________________________________________
Wekalist mailing list
Wekalist@list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

[prev in list] [next in list] [prev in thread] [next in thread]