WRRnnls: Special weighted and ridge penalized regression

In olangsrud/experimentalRpackage: An experimental R package containing experimental/early R functions

Description

By using nnls, ginv or glmnet

Usage

WRRnnls(x, y, xExact = NULL, yExact = NULL, wExact = 1000, lambda = 1e-04^2)

WRRginv(x, y, xExact = NULL, yExact = NULL, wExact = 1000, lambda = 1e-04^2)

WRRglmnet(
  x,
  y,
  xExact = NULL,
  yExact = NULL,
  wExact = 1000,
  lambda = exp(((2 * 12):(-17 * 2))/2),
  intercept = FALSE,
  standardize = FALSE,
  thresh = 1e-10,
  lower.limits = 0,
  ...
)

Arguments

x	Input matrix, each row is an ordinary observation
y	Ordinary response observation (vector or matrix)
xExact	Input matrix, each row is a highly weighted observation
yExact	Highly weighted response observation (vector or matrix)
wExact	Weight for highly weighted observations
lambda	Ridge regression penalty parameter (sequence when glmnet)
intercept	glmnet parameter
standardize	glmnet parameter
thresh	glmnet parameter
lower.limits	glmnet parameter
...	Further glmnet parameters

Value

Output from nnls, glmnet or coefficient estimate calculated using ginv

Author(s)

Øyvind Langsrud

Examples

x <- cbind(1:11, -11:-1, c(1, 1, 2, 2, 1, 1, 2, 2, 1, 1, 2))
y <- matrix(5 - sin(2 * (1:11)), dimnames = list(NULL, "y"))
x1 <- x[1:9, ]
x2 <- x[10:11, ]
y1 <- y[1:9, , drop = FALSE]
y2 <- y[10:11, , drop = FALSE]


# Generalized inverse estimation
ginvCoef <- WRRginv(x1, y1, x2, y2)

# Non-negative estimation by nnls 
# Coefficients in output element x 
nnlsCoef <- WRRnnls(x1, y1, x2, y2)$x

# Non-negative estimation by glmnet
# Take out best fit from matrix of coefficients 
gn <- WRRglmnet(x1, y1, x2, y2)
glmnetCoef <- coef(gn)[-1, which.max(gn$dev.ratio), drop = FALSE]

# Another estimation by glmnet (not non-negative)
# Take out best fit from matrix of coefficients
gnInf <- WRRglmnet(x1, y1, x2, y2, lower.limits = -Inf)
glmnetCoefInf <- coef(gnInf)[-1, which.max(gn$dev.ratio), drop = FALSE]

# All coefficients
coef4 <- as.matrix(cbind(ginvCoef, nnlsCoef, glmnetCoef, glmnetCoefInf))
colnames(coef4) <- c("ginv", "nnls", "glmnet", "glmnetInf")
print(coef4)

# Original y and fitted values. Close fit for last two observation.
cbind(y, x %*% coef4)

olangsrud/experimentalRpackage documentation built on Sept. 20, 2021, 10:32 p.m.