sera: Squared Error-Relevance Area (SERA)

In nunompmoniz/IRon: Solving Imbalanced Regression Tasks

Squared Error-Relevance Area (SERA)

Description

Computes an approximation of the area under the curve described by squared error of predictions for a sequence of subsets with increasing relevance

Usage

sera(
  trues,
  preds,
  phi.trues = NULL,
  ph = NULL,
  pl = FALSE,
  m.name = "Model",
  step = 0.001,
  return.err = FALSE,
  norm = FALSE
)

Arguments

trues	Target values from a test set of a given data set. Should be a vector and have the same size as the variable preds
preds	Predicted values given a certain test set of a given data set. Should be a vector and have the same size as the variable preds
phi.trues	Relevance of the values in the parameter trues. Use ??phi() for more information. Defaults to NULL
ph	The relevance function providing the data points where the pairs of values-relevance are known. Default is NULL
pl	Boolean to indicate if an illustration of the curve should be provided. Default is FALSE
m.name	Name of the model to be appended in the plot title
step	Relevance intervals between 0 (min) and 1 (max). Default 0.001
return.err	Boolean to indicate if the errors at each subset of increasing relevance should be returned. Default is FALSE
norm	Normalize the SERA values for internal optimisation only (TRUE/FALSE)

Value

Value for the area under the relevance-squared error curve (SERA)

Examples

library(IRon)
library(rpart)

if(requireNamespace("rpart")) {

   #' data(accel)

   form <- acceleration ~ .

   ind <- sample(1:nrow(accel),0.75*nrow(accel))

   train <- accel[ind,]
   test <- accel[-ind,]

   ph <- phi.control(accel$acceleration)

   m <- rpart::rpart(form, train)
   preds <- as.vector(predict(m,test))

   trues <- test$acceleration
   phi.trues <- phi(test$acceleration,ph)

   sera(trues,preds,phi.trues)
   sera(trues,preds,phi.trues,pl=TRUE, m.name="Regression Trees")
   sera(trues,preds,phi.trues,pl=TRUE, return.err=TRUE)

}

nunompmoniz/IRon documentation built on April 24, 2023, 1:20 p.m.