phew, explaining this would be equivalent to explaining the whole idea of kernel based learning. Instead of replicating everything which was said and written at least a hundred times, I will give you some hints:
- let's say the data cannot be divided by a linear hyperplane in its input space but by a polynomial. Then the linear hyperplane will cause a lot of errors and hence support vectors - the same applies for other kernel functions or inappropriate kernel parameters
If you have any difficulties in understanding those two hints I would suggest to learn much more about the way support vector machines work, e.g. at http://www.kernel-machines.org
How to load processes in XML from the forum into RapidMiner: Read this!