Hi,

I think it depends on your task. Do you need to normalize data for classification?

I such a case, I suggest that the normalization is done respect to the rows, I mean, you should normalize each sample independently from its class. Something like this:

Data(i,:) = Data(i,:) /sum(Data(i,:) );

This because:

- If you normalize a column, what happens if you have a new sample?

- If you consider all the data, what happens if you find more data?

- If you train a model on data normalized by class, how are you going to normalize a new test sample? 

Dear Giorgio Roffo,

Yes I am using it for classification problem, your suggestion is really interesting. Thank you. 

You are welcome :)

Thank you Mr.  Andreas. However, Giorgio suggested normalization with respect to sample, you are suggesting the other way.  My doubt is that , although aim of normalization is to make data values in the range of [0,1]  is there any specific way of normalization or depends on problem. 

Dear Giorgio Roffo, I was tried your suggestion but results became worse. Later, tried the following,  Got good results. Is this really acceptable.  

Is it possible to do feature normalization with respective to class. EX: 10x10 data matrix with two class. Each class of size 5x5. Now normalize 25 features of class 1 and 25 features of class 2 separately. Is this process acceptable.

Hi!

I also agree with Giorgio Roffo ,for  different task data, the normalization may be needed or not,  no matter how you are going to normalize the data(as sample , column or class) the defference between  features will be weaken,  which may influence your classification result 

I suggest   before  you normalize as column,  feature selection  would be necessary.

Thank you Mr. Li Yansong

For my point of view, only 3-th approach is good. To normalize each class is wrong, because it may be for first class and for second class values (after normalization) will be same, but really they are very different. To use second approach isn't good, because different features may have different measure (e.g., kg and m^2). Only third approach is good!

Hello. As others have already pointed out, you should normalise the features, since they may represent different physical variables. You are doing classification, therefore the statistics of the features are important. Normalisation implies that the normalised samples have a normal distribution, usually N(0,1), under the null (signal absent) hypothesis. This may or may not be possible depending on how much you know about the features' statistics and whether they are independent or not. There is a lot of history on this type of problem, particularly for radar and sonar, where this type of signal processing has been in use since the 1970s (look for CFAR detection - constant false alarm rate).