Salam Sir Alomari,

The Libsvm package proposes a matlab implementation for scaling data (link 1): (I haven't tested yet)

But, I do not see the added value of using the wrapper approach in the second stage because the best features have already been selected by the filter method.

In my opinion, You can either use libsvm as just a classifier  or integrate MRMR into the wrapper approach as in the attached file (in the second case, libsvm has to be integrated into the MRMR code).

May Allah help you.

Best regards.

http://stackoverflow.com/questions/10055396/scaling-the-testing-data-for-libsvm-matlab-implementation

Walikum Salam Dr malek

I have used hybrid approaches as in first stage i  used MRMR to select the best 100 features.Then, I used GA-SVM to find the goodness feature subset.

I confused  in which data format should be without negative affect or mislead the classifier?

A-  Should I run MRMR on scaling data then run svm on same data format??

 Or

B-    run MRMR on discretization data as suggested by author of MRMR then back to     convert original data scaling format (0-1) ??

Is it necessary to keep same data format in feature selection and classification process?

Thanks For Cooperation

Thank you for this explanation.

I do not have much experience to suggest the best one. But, in my opinion, if the feature selection (FS) approach MRMR do not depends on the classifier (SVM), there is not a need to use the same format because the aim of FS is to optimize the MRMR criterion. The same format can be useful if we integrate SVM  into MRMR as it is done in some papers.

Furthermore, you can test both of them and choose the one which gives the best classification accuracy (which is the main purpose of FS)

Hi Osama,

mRMR is a filter approach based on mutual information and GA-SVM is wrapper approach that used the classification accuracy rate as objective function.

The best way is to run mRmR first and selection the "K" best features. The next steps of your approach is used GA-SVM and select the relevant subset among the "K" best features.

To use GA-SVM, we must define a new binary variable X which has K feature and takes 0 is the feature is not selected and 1 if the feature is selected.

You can read my publication, i have used SA-SVM "Simulated annealing for feature selection"

http://www.mecs-press.org/ijitcs/ijitcs-v7-n5/IJITCS-V7-N5-1.pdf

 Thanks Dr Seyyid

what that data format you used to Run MRMR on it, then how about data format on wrapper approach>\??

So, My question is the changing of data format  cause information loss??

MRMR run on discretization data.

SVM accept scaling data between 0-1

Thanks for Cooperation

I do not understand why the data exchange ?

suppose that you have 5 features X = {X1, X2, X3, X4, X5}, you split the data into tow sets: learning set, test set.

you run mRmR on learning set and you pick the best features. suppose that the best features are : X1, X3, X4, X5

now, you run GA-SVM on X1, X3, X4, X5, in this stage :

the GA algorithm generate binary vector {1,0,0,1}

the SVM use the features : X1 and X2 to build the classification model.

I understand all process of feature selection Whether it in filter approach and wrapper approach.

I will tell why data should transform..

If you go through the literature you will find MRMR very efficent when the data is discretize..

So i discretize data and fit into MRMR, then MRMR reveals with best 100 features. NOW, LIBSVM stated by the author of it that" data should be in scaling data format".

Here, I am confused did the transformation of the data lead to loss of information. Since, the main two process run into different data format???!!!

I hope you get my question.