Machine learning techniques prove to be interesting and useful for many different problems, classification specially is a task where given a set of features a sample X must be labelled into a label Y
To do so a model must learn the conditional distribution P(Y|X) from the data, however when the number of available samples is small the learned models usually lack generalization capability
Given this here we present AugmenterR, the package is founded on a method based on conditional probability itself to generate novel samples, in the course of this document we present the two main functions an user will be interested in.
Generate
This function should mostly be used in scenarios where the interest of the user is mostly for regression tasks, as its normal use is as a step for our main function for creating data for classification. It has the following parameters
data Dataframe with the data we want to extend
regression: Default value false returns a list containing a novel sample as well as some information that the method uses to revalidate a sample, If you desire to generate samples who are not conditioned to a class for regression use regression=TRUE
the following is an example of how to use it
require(AugmenterR)
#> Loading required package: AugmenterR
NovelSample=Generate(iris,regression=TRUE)
print(NovelSample)
#> V1 V2 V3 V4 V5
#> X1.ncol.data. 4.373172 2.663216 NA NA <NA>
Here we see NovelSample as a sample generated from the iris dataset which respects its distribution, we present these properties later using the function that conditions on a class
GenerateMultipleCandidates:
This function creates many novel samples by conditioning then on a class of interest, this can be useful for both working with imbalanced datasets as we can balance the classes generating novel samples as well as augmenting all classes on small data problems. We present how to run the function as well as present how the novel samples compare to the old ones
require(AugmenterR)
Setosa=GenerateMultipleCandidates(iris,'setosa',5,0.9,40)
Virginica=GenerateMultipleCandidates(iris,'virginica',5,0.9,40)
Versicolor=GenerateMultipleCandidates(iris,'versicolor',5,0.9,40)
head(Setosa)
#> Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> X1.ncol.data. 4.974765 3.427988 1.332050 0.3 setosa
#> X1.ncol.data.1 5.020657 3.462028 1.525627 0.2 setosa
#> X1.ncol.data.2 4.289510 2.905976 1.299071 0.2 setosa
#> X1.ncol.data.3 4.606467 3.292332 1.228838 0.2 setosa
#> X1.ncol.data.4 5.109861 3.346060 1.340898 0.2 setosa
#> X1.ncol.data.5 4.427231 2.837835 1.336770 0.2 setosa
To present how the synthetic samples and novel samples come from the same distribution we show some comparisons between then: First histograms
df=data.frame(iris,source='Original')
Setosa=data.frame(Setosa,source='Sinthetic')
Virginica=data.frame(Virginica,source='Sinthetic')
Versicolor=data.frame(Versicolor,source='Sinthetic')
df=rbind(df,Setosa,Virginica,Versicolor)
require(ggplot2)
#> Loading required package: ggplot2
ggplot2::ggplot(df) + aes(x=Sepal.Length,col=source) + facet_wrap(~Species) + geom_histogram(aes(y=..density..) )
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Using PCA
pc=prcomp(df[,1:4],center=TRUE,scale=TRUE)$x
pc=data.frame(pc,df[,5:6])
ggplot2::ggplot(pc) + aes(x=pc[,1],y=pc[,2],col=Species) +facet_wrap(~source) + labs(x='First Component',y='Second Component') + geom_point()
