Hi,

It is better to test the impact of the data normalization as a pre-processing step in any data mining task. The format [-1,1] is just used to refer to the label of each instance and does not related to the feature vectors. In a standard normalization, all the features are scaled to the range of [0,1] by a common normalization technique (see the below link for instance).You should also be noticed that the parameters of SVM like as c and gamma are required to be tuned via, e.g., a greed search or simply a trial and error technique.

Useful link: 

https://www.researchgate.net/post/How_to_normalize_the_feature_vectors_for_SVM_training/1

salem,

try to transform your problem to multiple binary classification problems, SVM will return better results (SVM performs binary classification).

In general, the preprocessing tasks, we should do depend on the classifier used in the traing step. with SVM, features should be numeric and normalised.

see the experimental part of this papers :

http://ieeexplore.ieee.org/document/7011237/

http://www.sciencedirect.com/science/article/pii/S1084804511000038

Thanks Hadi Shakiban and Wassila Guendouzi. I have used a common technique for both testing and training set. And I am using k- fold cross validation to choose the best parameters. And I got in the libsvm guide that we need to convert the data to svm format.and svm format takes only numeric data so I have scaled to [-1,1] before converting to svm format. should I try using [0,1]? will it be helpful? 

Nabasmita,

For better data quality, at perpeocessing stage, you can consider removing the noise from the data. On the other hand, if you have a high-dimensional dataset you may consider attribute selection method to in order to enhance classification results. 

Regarding the SVM algorithm, you need to know first how the tool/code (that you're using) implements such algorithm. Some tools apply John Platt's SVM that *by default* transforms nominal attributes into binary, from the one hand, and normalizes all attributes, from the other hand.

Fianlly, you need to figure out your normalization range based on your task.

HTH.

Samer

It is no need to do normalization when the classifier is Neural network.

But according to these papers,  we can use the normalized features and preprocessing steps  to get the better accuracy.

http://www.ieeeexplore.ws/document/7222636/

http://www.ieeeexplore.ws/document/7047636/

Thanks

Thankyou everyone. I performed the 10 class classification. The normalized data is in the range [-1,1] and I havenot used the svm format.And  for 2 subjects I got appreciable recognition rates of around 93%. But for another one subject i got very low acuraccy. I think this is because there is very less variation in the datas of the classes.