ncordon / imbalance Goto Github PK

Home Page: https://ncordon.github.io/imbalance

License: GNU General Public License v2.0

R 86.12% C++ 13.88%

oversampling binary-classification imbalanced-data r

imbalance's Introduction

imbalance

imbalance provides a set of tools to work with imbalanced datasets: novel oversampling algorithms, filtering of instances and evaluation of synthetic instances.

Installation

You can install imbalance from Github with:

# install.packages("devtools")
devtools::install_github("ncordon/imbalance")

Examples

Run pdfos algorithm on newthyroid1 imbalanced dataset and plot a comparison between attributes.

library("imbalance")
data(newthyroid1)

newSamples <- pdfos(newthyroid1, numInstances = 80)
# Join new samples with old imbalanced dataset
newDataset <- rbind(newthyroid1, newSamples)
# Plot a visual comparison between both datasets
plotComparison(newthyroid1, newDataset, attrs = names(newthyroid1)[1:3], cols = 2, classAttr = "Class")

After filtering examples with neater:

filteredSamples <- neater(newthyroid1, newSamples, iterations = 500)
#> [1] "12 samples filtered by NEATER"
filteredNewDataset <- rbind(newthyroid1, filteredSamples)
plotComparison(newthyroid1, filteredNewDataset, attrs = names(newthyroid1)[1:3])

Execute method ADASYN using the wrapper provided by the package, comparing imbalance ratios of the dataset before and after oversampling:

imbalanceRatio(glass0)
#> [1] 0.4861111
newDataset <- oversample(glass0, method = "ADASYN")
imbalanceRatio(newDataset)
#> [1] 0.9722222

imbalance's People

Contributors

Stargazers

Watchers

Forkers

zhaoxiaohe guhjy zawaruddin angelina1996 minghao2016

imbalance's Issues

bnlearn gives an error when a factor has only one level

Checking with cars data:

 Error in check.data(x, allowed.types = c(discrete.data.types, continuous.data.types)) : 
  variable safety must have at least two levels.

rwo() does not support non-numeric factors properly

The documentation of the rwo function claims it can handle every type of dataset. Indeed, the 2nd part of the method adds noise to numeric features and samples from the existing values for all other types of features. However, it can fail earlier because it attempts to convert all features/columns to numeric (which IMO is not necessary in this algorithm, but probably just copy-pasted, as it also exists at the beginning of the other oversampling methods [and makes sense there]). For example, if we have a factor which has non-numeric levels, rwo throws an error when executing dataset <- toNumeric(dataset, exclude = classAttr).

# Does work
imbalance::rwo(data.frame(test1 = rnorm(10), test2 = rnorm(10), class = factor(sample(c("a", "b"), 10, T))), numInstances = 5, classAttr = "class")
# Error
imbalance::rwo(data.frame(test1 = rnorm(10), test2 = factor(sample(c("a", "b"), 10, T)), class = factor(sample(c("a", "b"), 10, T))), numInstances = 5, classAttr = "class")

oversample() gives an error if ratio >= 0.98 for SMOTE

Sample code below to illustrate the issue. Documentation says that ratio can be between 0 and 1. However, oversample() gives an error if the ratio specified is really close to 1, and also error for ratio = 1. How can a user work around this?

newDataset <- oversample(glass0, ratio = 0.9, method = "SMOTE") #No problem
newDataset <- oversample(glass0, ratio = 1, method = "SMOTE")
Error in sample.int(length(x), size, replace, prob) :
cannot take a sample larger than the population when 'replace = FALSE'
newDataset <- oversample(glass0, ratio = 1.0, method = "SMOTE")
Error in sample.int(length(x), size, replace, prob) :
cannot take a sample larger than the population when 'replace = FALSE'
newDataset <- oversample(glass0, ratio = 0.95, method = "SMOTE")
newDataset <- oversample(glass0, ratio = 0.99, method = "SMOTE")
Error in sample.int(length(x), size, replace, prob) :
cannot take a sample larger than the population when 'replace = FALSE'
newDataset <- oversample(glass0, ratio = 0.98, method = "SMOTE")
Error in sample.int(length(x), size, replace, prob) :
cannot take a sample larger than the population when 'replace = FALSE'
newDataset <- oversample(glass0, ratio = 0.97, method = "SMOTE") #No error for ratio = 0.97

Improve bandwidth parameter search in PDFOS

Currently a naive mechanism is being used, provided that we know that best bandwidth is going to be O(Silverman's rule of thumb).

  for(double v = 0.25; v < 1.5; v = v + 0.05){
    possible_bwidth.push_back(v * silverman_bandwidth);
  }

Find a better way to adapt the parameter. Maybe a simulated annealing (?).

sqrt(int)

Documented in https://cran.r-project.org/doc/manuals/r-patched/R-exts.html#Portable-C-and-C_002b_002b-code