Learning Metrics via Discriminant Kernels and Multidimensional Scaling: Toward Expected Euclidean Representation