Orange Forum • View topic - Feature selection techniques in Orange

Feature selection techniques in Orange

General discussions about Orange and with Orange connected things (data mining, machine learning, bioinformatics...).

Feature selection techniques in Orange

Postby dimich » Mon Mar 28, 2011 18:43

Hello,

FIrstly, I would like to express my excitement about Orange tool and its usability.

What are the measures that Orange supports for quality of attributes? I think it focuses maily on filter selection techniques, as reported at "Measures for Classification Problems" section in this link -> http://orange.biolab.si/doc/reference/M ... ribute.htm. Is it interesting for you the creation of additional feature selection techniques for measuring the quality of attributes? There are other categories also containing wrapper and embedded methods.

Moreover, one can find filter methods divided into more categories like univariate and multivariate.
-> http://bioinformatics.oxfordjournals.or ... l.pdf+html

Are you interested in such an extension?

Re: Feature selection techniques in Orange

Postby Blaz » Mon Mar 28, 2011 20:42

You are right. Orange has indeed a number of methods for filter-based methods, and does not cover well wrapper based approaches. I believe there are some (or at least one) wrapper in orngFSS, but that one is quite trivial. There is no systematic coverage of a number of different approaches, like the ones mentioned in Saeys et al (2007) that you have nicely pointed out. Also, I have always wished for an close-to-exhaustive search for a small sets of features that one could run in case of rather smaller data sets.

We've been interested in FSS since Orange has started, and perhaps still need to be convinced if computationally heavy approaches pay off. For instance, does wrapper-based approach help SVM and random forests to perform better? So, an idea is to (in your project) propose a library of FSS approaches (e.g., an extension of orngFSS), implement a set of examples and, perhaps most important, a demonstration to show which techniques work well. I think this could be nice. Go ahead and make a good proposal!

Best wishes,
Blaz.

Re: Feature selection techniques in Orange

Postby dimich » Tue Mar 29, 2011 15:01

Ok Blaz, thanks for the answer!


Return to General Discussions