Exploiting sparsity to build efficient kernel based collaborative filtering for top-N item recommendation