weighted.one.vs.rest {costsensitive}

Weighted One-Vs-Rest

Description

Creates a cost-sensitive classifier by creating one classifier per class to predict cost. Takes as input a classifier accepting observation weights. The objective is to create a model that would predict the class with the minimum cost.

Usage

weighted.one.vs.rest(X, C, classifier, predict_type_prob = "prob",
  wap_weights = FALSE, nthreads = 1, ...)

Arguments

X

The data (covariates/features).

C

matrix(n_samples, n_classes) Costs for each class for each observation.

classifier

function(X, y, weights=w, ...) -> object, that would create a classifier with method 'predict'. The 'y' vector passed to it is of class 'integer' with values 0/1 only.

predict_type_prob

argument to pass to method 'predict' from the classifier passed to 'classifier' in order to output probabilities or numeric scores instead of classes (i.e. 'predict(object, newdata, type=predict_type_prob')).

wap_weights

Whether to use the weighting technique from the 'Weighted-All-Pairs' algorithm.

nthreads

Number of parallel threads to use (not available on Windows systems). Note that, unlike the Python version, this is not a shared memory model and each additional thread will require more memory from the system. Not recommended to use when the algorithm is itself parallelized.

...

Extra arguments to pass to 'classifier'.

References

Beygelzimer, A., Dani, V., Hayes, T., Langford, J., & Zadrozny, B. (2005, August). Error limiting reductions between classification tasks.

Examples

library(costsensitive)
wrapped.logistic <- function(X, y, weights, ...) {
	return(glm(y ~ ., data = X, weights = weights, family = "quasibinomial", ...))
}
set.seed(1)
X <- data.frame(feature1 = rnorm(100), feature2 = rnorm(100), feature3 = runif(100))
C <- data.frame(cost1 = rgamma(100, 1), cost2 = rgamma(100, 1), cost3 = rgamma(100, 1))
model <- weighted.one.vs.rest(X, C, wrapped.logistic, predict_type_prob = "response")
predict(model, X, type = "class")
predict(model, X, type = "score")
print(model)