Skip to contents

RUSBoost for two-class problems

Source: R/rusboost.R
rusboost.Rd

RUSBoost for two-class problems.

Usage

rusboost(formula, df, size, ir = 1, learn_rate = 1, rus = TRUE, control)

# S3 method for class 'rusboost'
predict(object, newdata, type = "prob", ...)

Arguments

formula

A formula specify predictors and target variable. Target variable should be a factor of 0 and 1. Predictors can be either numerical and categorical.

df

A df frame used for training the model, i.e. training set.

size

Positive integer giving the maximum number of weak-learner attempts.

ir

Imbalance ratio. Specifies how many times the under-sampled majority instances are over minority instances.

learn_rate

Positive finite learning rate. Default of 1.

rus

TRUE for random undersampling; FALSE for AdaBoost with full sample

control

Control object passed onto rpart function.

object

A fitted rusboost object.

newdata

A data frame containing the predictors used in training.

type

Prediction type: "prob" for the weighted share of positive votes, or "class" for the sign of the weighted class vote.

...

Additional arguments; currently unused.

Value

rusboost() returns an object of class rusboost, a list containing fitted trees (weakLearners), their voting weights (alpha), observation-weight vectors (w), and the training formula. predict() returns a numeric vector of weighted positive-vote shares for type = "prob", or classes -1 and 1 (0 for a tied vote) for type = "class".

Details

A learner that correctly classifies every observation in the full training data stops training successfully. It replaces any earlier learners and is returned as the sole predictive learner with voting weight 1, a finite representation of its limiting vote. Its observation-weight vector is the normalized vector used for that attempt.

Learners with weighted error at least one half receive no vote. AdaBoost (rus = FALSE) stops at such a learner; RUSBoost skips it and tries another sample, up to size attempts. Training raises an error if no usable learner is found. The response must contain two nonmissing classes, including class 1.

Examples

training <- data.frame(x = 1:8, y = factor(c(rep(0, 4), rep(1, 4))))
control <- rpart::rpart.control(minsplit = 2, minbucket = 1, cp = 0, xval = 0)
fit <- rusboost(y ~ x, training, size = 3, rus = FALSE, control = control)
predict(fit, training, type = "prob")
#> [1] 0 0 0 0 1 1 1 1
predict(fit, training, type = "class")
#> [1] -1 -1 -1 -1  1  1  1  1