> SVM is known to defeat the curse of dimensionality...then why is it so
> that having a large number of attributes/features in the data is not
> desirable when using SVM?
> I thought it should not matter....where am I getting confused?
that depends on the kernel you are using. Kernels involving skalarproducts
as the linear kernel, or polynomial kernels, have no problems
with many features: if a feature is not significant, the corresponding
entry in the weightvector w will be zero.
But if you use a kernel which involves ||x-y|| as in the radial kernel,
features which contain no information will lead to unnecessary large
values and get artificially signifacant for your machine.
What works in some situations is to try linear classification first,
then study the weight vector w and remove features corresponding to
zero entries in w. After this step you can use other kernels, I
start with a radial kernel in most situations.