Learning the kernel matrix in discriminant analysis via quadratically constrained quadratic programming