Learning kernels for variants of normalized cuts: Convex relaxations and applications