HPLBmatrix {HPLB}

Pairwise Total Variation Distance Lower Bound Matrix for the Multi-Class Setting

Description

Pairwise Total Variation Distance Lower Bound Matrix for the Multi-Class Setting

Usage

HPLBmatrix(
  labels,
  ordering.array,
  alpha = 0.05,
  computation.type = "non-optimized",
  seed = 0,
  ...
)

Arguments

labels

a numeric vector value. The labels of the classes, should be encoded in [0,nclass-1].

ordering.array

a numeric array of size (nclass, nclass, nobs) such that the value (i,j,k) represents a propensity of being of class j instead of i for observation k.

alpha

a numeric value. The type-I error level.

computation.type

a character value. For the moment only "non-optimized" (default) available.

seed

an integer value. The seed for reproducility.

...

additional parameters to be passed to the HPLB function.

Value

a numeric matrix of size (nclass, nclass) giving the matrix of pairwise total variation lower bounds.

Author(s)

Loris Michel, Jeffrey Naef

References

L. Michel, J. Naef and N. Meinshausen (2020). High-Probability Lower Bounds for the Total Variation Distance

Examples

 # iris example
 require(HPLB)
 require(ranger)

 # training a multi-class classifier on iris and getting tv lower bounds between classes
 data("iris")

 ind.train <- sample(1:nrow(iris), size = nrow(iris)/2, replace = FALSE)

 rf <- ranger(Species~., data = iris[ind.train, ], probability = TRUE)
 preds <- predict(rf, iris[-ind.train,])$predictions

 # creating the ordering array based on prediction differences
 ar <- array(dim = c(3, 3, nrow(preds)))
 for (i in 1:3) {
   for (j in 1:3) {
    ar[i,j,] <- preds[,j] - preds[,i]
   }
 }

 # encoding the class response
 y <- factor(iris$Species)
 levels(y) <- c(0,1,2)
 y <- as.numeric(y)-1

 # getting the lower bound matrix
 tvhat.iris <- HPLBmatrix(labels = y[-ind.train], ordering.array = ar)
 tvhat.iris