useR! 2024

Random projection is a powerful and important tool for dimensionality reduction where a set of high-dimensional points is linearly mapped onto a lower dimension. Random projection matrices can be rapidly generated and are oblivious to the data distribution, maintain interpretability and are equipped with theoretical guarantees on preserving the geometry of the original space with a high probability. When employed in a supervised setting, they can provide a significant reduction in computational cost. However, they tend to overfit so it is desirable to first eliminate the unimportant predictors and then perform the random projection and estimate the model on the space of the reduced (i.e., projected) predictors. Moreover, to reduce the uncertainty from the random projection, ensembles can be built. In this work we propose an R package which implements a variety of random projection and screening tools for regression in high-dimensional settings. The functionality of the package is presented using simulated and real data examples.