The Generalized Singular Value Decomposition

GSVD

The generalized singular value decomposition (GSVD) is a name shared by two different SVD techniques. This package is for the “weighted” or “vector-constrained” GSVD. For one of the most straight forward introductions to the SVD and GSVD see the Appendices of Greenacre (1984).

Background

The GSVD generalizes in two ways:

it generalizes the standard SVD to allow for constraints or weights to be applied to the left and right singular vectors, and
it generalizes many multivariate techniques, e.g., principal components analysis (PCA), correspondence analysis (CA), canonical correlation analysis (CCA), partial least squares (PLS), multidimensional scaling (MDS), linear discriminant analysis (LDA), and many many more.

The GSVD is an extraordinarily powerful and flexible tool for multivariate analyses and it is the core technique in the French school of data science/analyses (Holmes & Josse, 2017).

Package overview

This GSVD package is the first and most important package in the family of ExPosition2 packages. For an exposition of ExPosition see Beaton et al., (2014). The GSVD package is a focused package with one goal: give R users better and simpler access to the GSVD. GSVD’s companion packages will allow users more direct access to specific methods.

gsvd() is an efficient pure R implementation of the GSVD with no (current) dependencies. However, in the not-so-distant future, we plan to make GSVD better, faster, and more efficient with the use of Rcpp and Matrix.

Installation

The GSVD package is not yet on CRAN but should be soon. For now, the simplest approach to installation is through the devtools package:

# devtools install via github
devtools::install_github("derekbeaton/GSVD")

Functions

Core

tolerance_eigen() is an alternative to the eigenvalue decomposition function to only return vectors and values above some precision threshold (e.g., .Machine$double.eps)
tolerance_svd() is an alternative to the SVD function to only return vectors and values above some precision threshold (e.g., .Machine$double.eps)
geigen() is the generalized eigen function. It passes through to tolerance_eigen().
gsvd() is the generalized SVD function. It passes through to tolerance_svd().
gplssvd() is the generalized partial least squares-singular value decomposition function. It passes through to tolerance_svd().

Bells-and-whistles

sqrt_psd_matrix() computes the square root of a square positive semi-definite (psd) matrix.
invsqrt_psd_matrix computes the inverse of the square root of a square positive semi-definite (psd) matrix
A small set of functions to check for specific types of square matrices. See utils.R.

Usage

The snippets below are very abbreviated examples. There are more—and more detailed—examples in vignettes as well as in the preprint for this package (Beaton, 2020).

One table analyses

Here we provide three examples of “one table” analyses: principal components analysis, multidimensional scaling (distances), and correspondence analysis (with a smidgen of multidimensional scaling). These make use of gsvd() and geigen()

library(GSVD)

# several examples of principal component analysis
 data(wine)
 wine.objective <- wine$objective
 ## "covariance" PCA
 cov.pca.data <- scale(wine.objective,scale=FALSE)
 cov.pca.res <- gsvd(cov.pca.data)
 ## "correlation" PCA
 cor.pca.data <- scale(wine.objective,scale=TRUE)
 cor.pca.res <- gsvd(cor.pca.data)
 ## an alternative approach to "correlation" PCA with GSVD constraints
 cor.pca.res2 <- gsvd(cov.pca.data,RW=1/apply(wine.objective,2,var))

# an example of multidimensional scaling
  D <- as.matrix(dist(wine$objective))^2
  masses <- rep(1/nrow(D), nrow(D))
  Xi <- matrix(-masses, length(masses), length(masses))
  diag(Xi) <- (1-masses)
  mds.res_geigen <- geigen((-D / (nrow(D) * 2)), Xi)

 
# an example of correspondence analysis.
 data(authors)
 Observed <- authors/sum(authors)
 row.w <- rowSums(Observed)
   row.W <- diag(1/row.w)
 col.w <- colSums(Observed)
   col.W <- diag(1/col.w)
 Expected <- row.w %o% col.w
 Deviations <- Observed - Expected
 ca.res <- gsvd(Deviations,row.W,col.W)
 
 
 # an alternate example of correspondence analysis by way of multidimensional scaling of Chi-squared distances
 Chi2DistanceMatrix <- t(Deviations) %*% diag(1/row.w) %*% Deviations
 ca.res_geigen <- geigen(Chi2DistanceMatrix, col.W)

Two table analyses

Here we provide four examples of “two table” analyses all through gplssvd(): partial least squares correlation, canonical correlation analysis, reduced rank regression/redundancy analysis, and partial least squares-correspondence analysis. This also requires data from the ExPosition package. Each of these techniques can be expressed as optimization of latent vectors.

library(GSVD)

  data(wine)
  X <- scale(wine$objective)
  Y <- scale(wine$subjective)
  
  ## an example of partial least squares-svd (aka PLS correlation)
  pls.res <- gplssvd(X, Y)
  
  pls.res$d
  diag( t(pls.res$lx) %*% pls.res$ly )
  
  ## Canonical correlation analysis (CCA)
  ### NOTE:
  #### This is not "traditional" CCA because of the generalized inverse.
  #### However the results are the same as standard CCA when data are not rank deficient.
  #### and this particular version uses tricks to minimize memory & computation
  cca.res <- gplssvd(
    X = MASS::ginv(t(X)),
    Y = MASS::ginv(t(Y)),
    XRW=crossprod(X),
    YRW=crossprod(Y),
    scale_X = F,
    scale_Y = F
  )
  cca.res$d
  diag( t(cca.res$lx) %*% cca.res$ly )  
  
  ## an example of reduced rank regression/redundancy analysis
  rrr.res <- gplssvd(X, Y, XRW=MASS::ginv(crossprod(X)))
    ## to note: rrr.res$fi is "beta" and rrr.res$v is alpha (see rrr.nonmiss: http://ftp.uni-bayreuth.de/math/statlib/S/rrr.s)
  rrr.res$d
  diag( t(rrr.res$lx) %*% rrr.res$ly )

  
  ## an example of pls-correspondence analysis (see https://utd.edu/~herve/abdi-bdAa2015_PLSCA.pdf)
  library(ExPosition)
  data("snps.druguse")
  X_nom <- makeNominalData(snps.druguse$DATA1)
    Ox <- X_nom / sum(X_nom)
    rx <- rowSums(Ox)
    cx <- colSums(Ox)
    Ex <- rx %o% cx
    Zx <- Ox - Ex
  Y_nom <- makeNominalData(snps.druguse$DATA2)
    Oy <- Y_nom / sum(Y_nom)
    ry <- rowSums(Oy)
    cy <- colSums(Oy)
    Ey <- ry %o% cy
    Zy <- Oy - Ey
  
  plsca.res <- gplssvd(Zx, Zy, 
                       XLW = 1/rx, YLW = 1/ry,
                       XRW = 1/cx, YRW = 1/cy)
  plsca.res$d
  diag(t(plsca.res$lx) %*% plsca.res$ly)

Et voila! We have a unified generalized framework for many standard multivariate analyses, all through the g*() family of functions here in the GSVD package. (To note: the above two table analyses could also have been done through the gsvd() but it is more convenient to do so with gplssvd().)

And beyond!

GSVD is the first package part of the larger ExPosition2 family. More to come soon!

References

Greenacre, M. (1984). Theory and applications of correspondence analysis. Academic Press.
Holmes, S., & Josse, J. (2017). Discussion of “50 Years of Data Science”. Journal of Computational and Graphical Statistics, 26(4), 768-769.
Beaton, D., Fatt, C. R. C., & Abdi, H. (2014). An ExPosition of multivariate analysis with the singular value decomposition in R. Computational Statistics & Data Analysis, 72, 176-189.
Beaton, D. (2020). Generalized eigen, singular value, and partial least squares decompositions: The GSVD package. ArXiv:2010.14734 [Cs, Stat]. http://arxiv.org/abs/2010.14734