unbalanced (version 2.0)

unbalanced-package: Racing for Unbalanced Methods Selection

Description

A dataset is said to be unbalanced when the class of interest (minority class) is much rarer than normal behaviour (majority class). The cost of missing a minority class is typically much higher that missing a majority class. Most learning systems are not prepared to cope with unbalanced data and several techniques have been proposed to rebalance the classes. This package implements some of most well-known techniques and propose a racing algorithm [2] to select adaptively the most appropriate strategy for a given unbalanced task [1].

Details

Package:

unbalanced

Type:

Package

Version:

2.0

Date:

2015-06-17

License:

GPL (>= 3)

References

1. Dal Pozzolo, Andrea, et al. "Racing for unbalanced methods selection." Intelligent Data Engineering and Automated Learning - IDEAL 2013. Springer Berlin Heidelberg, 2013. 24-31. 2. Birattari, Mauro, et al. "A Racing Algorithm for Configuring Metaheuristics."GECCO. Vol. 2. 2002.

See Also

ubBalance, ubRacing

Examples

#use Racing to select the best technique for an unbalanced dataset
library(unbalanced)
data(ubIonosphere)

#configure sampling parameters
ubConf <- list(type="ubUnder", percOver=200, percUnder=200, k=2, perc=50, method="percPos", w=NULL)

#load the classification algorithm that you intend to use inside the Race
#see 'mlr' package for supported algorithms
library(randomForest)
#use only 5 trees
results <- ubRacing(Class ~., ubIonosphere, "randomForest", positive=1, ubConf=ubConf, ntree=5)

# try with 500 trees
# results <- ubRacing(Class ~., ubIonosphere, "randomForest", positive=1, ubConf=ubConf, ntree=500)
# let's try with a different algorithm
# library(e1071)
# results <- ubRacing(Class ~., ubIonosphere, "svm", positive=1, ubConf=ubConf)
# library(rpart)
# results <- ubRacing(Class ~., ubIonosphere, "rpart", positive=1, ubConf=ubConf)