When dealing with massive datasets, how do you decide between using sparse matrix representations versus dense ones, and how does this affect memory bandwidth?