I think there can be a simple explanation of why data needs to be scaled. The reason is that SVM training tries to maximize the distance between the training samples of two different classes. If all the dimensions are not at the same scale, the difference of two training samples can be dominated by a dimension that is very large compared to other dimensions, even though the other dimensions may be more discriminative.
- Amit
On Thu, 3 Jun 2004, Vojislav Kecman wrote:
> Hi,
>
> > How does scaling of the data relate to kernel selection? Can they be
> > seen as equivalent in some way?
>
> It's not related to the type of the kernel, but it is a MUST if you want to
> have good results. Just scale the inputs between -1 and 1, or to the data
> having zero mean and variance = 1.
> Don't ask me why! It is longer than average Email letter.
> Just do it, and if curious, search the papers and books for the answer.
>
> Regards,
>
> VK
> --
> *...........*...........*...........*...........*...........*...........*
> Vojislav Kecman, Assoc. Prof.
> School of Engineering
> The University of Auckland
> Private Bag 92019
> Auckland, New Zealand
>
> Web site: http://www.support-vector.ws
> or: http://www.engineers.auckland.ac.nz/~vkec001/
> E-Mail: [log in to unmask]
>
> Phone: + 64 9 37 37 599 Ext. 88178
> Fax: + 64 9 37 37 479
>
---
The meme for blind faith secures its own perpetuation by the simple unconscious expedient of discouraging rational inquiry. - Richard Dawkins
|