undersampling: Undersampling

In dalpozz/warping: Probability warping and calibration with undersampling

Description

Undersampling the majority class with rate = beta, where beta is the probability of selecting a negative instance during undersampling. The probability estimate is calibrated to adjust it for class prior change

Usage

undersampling(task, lrn, beta, metrics, rdesc = NULL, positive, negative,
  verbose = TRUE, dirPlot = NA, ncore = 1, type = "within", B = 1)

Arguments

task	classification task to perform (see mlr package)
lrn	learner to use with the task (see mlr package)
beta	probability of selecting a negative instance during undersampling
metrics	metrics to asess classification performances
rdesc	resampling methods to use (see mlr package)
positive	value of the positive (minority) class
negative	value of the negative (majority) class
verbose	print extra information (logical variable)
dirPlot	directory where to save plots (set dirPlot=NA to avoid plots)
ncore	number of cores to use in multicore computation
type	way to apply undersampling: within or before the CV
B	number of models to create for each fold of the CV

Value

The function returns a list:

under	results with undersampling
cal	results after calibration
pred.under	predictions with undersampling
pred.cal	predictions after calibration
prob	posterior probability of the classifier for the positive class

Examples

library(mlbench)
data(Ionosphere)
library(mlr)
task <- makeClassifTask(data=Ionosphere, target="Class", positive="bad")
library(randomForest)
rf <- makeLearner("classif.randomForest", predict.type = "prob")
metrics <- list(f1, auc)
library(warping)
res <- undersampling(task, rf, beta=0.6, metrics, positive="bad", negative="good")

dalpozz/warping documentation built on May 14, 2019, 3:32 p.m.