Learning to rank using 1-norm regularization and convex hull reduction