Abstract:
|
A novel idea is presented for graphically investigating certain aspects of the geometry of high-dimensional data and, more generally, of Hilbert space-valued data. The method can be viewed as projecting the data, in a non-linear way, onto two-dimensional planes, and, given a data set of size n, the method constructs n(n-1)/2 different such projections. The usefulness of the methodology is illustrated, among others, by investigating its use in selecting a kernel for an SVM classifier.
|