Sparse orthonormalized partial least squares
Marcel A. J. van Gerven, Tom Heskes · 2010
Orthonormalized partial least squares (OPLS) is often used to find a low-rank mapping between inputs X and outputs Y by estimating loading matrices A and B. In this paper, we introduce sparse orthonormalized PLS as an extension of conventional PLS that finds sparse estimates of A through the use of the elastic net algorithm. We apply sparse OPLS to the reconstruction of presented images from BOLD response in primary visual cortex. Sparse OPLS finds solutions with low reconstruction error which are easy to interpret due to the sparseness of the loading matrix B. Moreover, the elastic net algorithm is generalized to allow for coupling constraints that induce a spatial regularization.