Exploring New Forms of Random Projections for Prediction and Dimensionality Reduction in Big-Data Regimes