fit_predictionBands: Fit conformal prediction bands based on density estimation...

In rizbicki/predictionBands:

Description

Fit conformal prediction bands based on density estimation for regression

Usage

fit_predictionBands(
  x,
  y,
  per_train = 0.4,
  per_val = 0.1,
  per_ths = 1 - per_train - per_val,
  k = max(round(per_ths * nrow(as.matrix(x))/100), 1),
  regressionFunction = FlexCoDE::regressionFunction.Forest,
  ...
)

Arguments

x	Matrix with covariates of training set
y	Vector with the (continuous) responses of training set
per_train	# percentage of samples used for traning density estimator (defaults to 40%)
per_val	# percentage of samples used for tuning density estimator (defaults to 10%)
per_ths	# percentage of samples used for computeing thresholds for the conformal method (defaults to 50%)
k	# Number of clusters for cd-split. Default to round(per_ths*nrow(as.matrix(x))/100) so that each cluster has on average 100 samples
regressionFunction	# regression function to be used for FlexCode. Defaults to Random Forests. See FlexCode documentation for additional regression methods.
...	Additional arguments to FlexCoDE::fitFlexCoDE

Value

Returns an object of the class predictionBands with the following components:

density_fit	Object of the class FlexCoDE with the estimated density
cum_dist_evaluated_train	Cumulative conditional distribution functions on the training set (for dist-split)
conformity_score_train	Conformal scores on the training set (for cd-split)
t_grid	Level sets of the densities
g_train	Profiles of the training sample
center_kmeans	The center of the clusters found by kmeans (in the profile space)

Examples

# generate data
n <- 1000
n_new <- 50
d <- 10
data <- matrix(NA,n,d+1)
x <- matrix(rnorm(n*d),n,d)
y <- x[,1]+rnorm(n,0,0.1)
fit <- fit_predictionBands(x,y,0.5,0.4,0.1)

xnew <- matrix(rnorm(n_new*d),n_new,d)
ynew <- xnew[,1]+rnorm(n_new,0,0.1)

 # Dist-split
 bands <- predict(fit,xnew,type="dist")
 bands[[1]]
 bands[[2]]
 plot(bands)
 plot(bands,ynew)

 # CD-split
 bands <- predict(fit,xnew,type="cd")
 bands[[1]]
 bands[[2]]
 plot(bands)
 plot(bands,ynew)

rizbicki/predictionBands documentation built on July 29, 2021, 2:01 a.m.