Projecting "better than randomly": How to reduce the dimensionality of very large datasets in a way that outperforms random projections