WPVI: p-Wasserstein Variable Importance

In WpProj: Linear p-Wasserstein Projections

p-Wasserstein Variable Importance

Description

This function will measure how much removing each covariate harms prediction accuracy.

Usage

WPVI(
  X,
  eta,
  theta,
  pred.fun = NULL,
  p = 2,
  ground_p = 2,
  transport.method = transport_options(),
  epsilon = 0.05,
  OTmaxit = 100,
  display.progress = FALSE,
  parallel = NULL
)

Arguments

X	Covariates
eta	Predictions from the estimated model
theta	Parameters from the estimated model.
pred.fun	A prediction function. must take variables x, theta as arguments: pred.fun(x,theta)
p	Power of Wasserstein distance
ground_p	Power of distance metric
transport.method	Transport methods. See transport_options() for more details.
epsilon	Hyperparameter for Sinkhorn iterations
OTmaxit	Maximum number of iterations for the Wasserstein method
display.progress	Display intermediate progress
parallel	a foreach backend if already created

Value

Returns an integer vector ranking covariate importance from most to least important.

Examples

n <- 128
p <- 10
s <- 99
x <- matrix(1, nrow = n, ncol = p )
beta <- (1:10)/10
y <- x %*% beta 
post_beta <- matrix(beta, nrow=p, ncol=s) 
post_mu <- x %*% post_beta

fit <-  WpProj(X=x, eta=post_mu, power = 2.0)
WPVI(X = x, eta = post_mu, theta = post_beta, transport.method = "hilbert")

WpProj documentation built on May 29, 2024, 7:55 a.m.