Ensembles of Random Projections for Nonlinear Dimensionality Reduction
Dimensionality reduction methods are widely used in information
processing systems to better understand the underlying structures
of datasets, and to improve the efficiency of algorithms for big
data applications. Methods such as linear random projections have
proven to be simple and highly efficient in this regard, however,
there is limited theoretical and experimental analysis for nonlinear
random projections. In this study, we review the theoretical framework
for random projections and nonlinear rectified random projections,
and introduce ensemble of nonlinear maximum random projections.
We empirically evaluate the embedding performance on 3
commonly used natural datasets and compare with linear random
projections and traditional techniques such as PCA, highlighting
the superior generalization performance and stable embedding of
the proposed method.