A Kernel Approach to Metric Multidimensional Scaling
The solution for the parameters of a nonlinear mapping in a metric multidimensional scaling by transformation, in which a stress criterion is optimised, satisfies a nonlinear eigenvector equation, which may be solved iteratively. This can be cast in a kernel-based framework in which the configuration of training samples in the transformation space may be found iteratively by successive linear projections, without the need for gradient calculations. A new data sample can be projected using knowledge of the kernel and the final configuration of data points.
Keywordsmultidimensional scaling kernel representation nonlinear feature extraction
Unable to display preview. Download preview PDF.