SMOTE sampling in caret package in R

Question

when using caret packge in the trainControl you can use "smote" sampling. what is the default parameters the train in caret are using for smote??

parameters such as: perc.over = 300, k = 8, perc.under = 100..

ctrl <- trainControl(method = "repeatedcv",
                     number = 10,
                     repeats = 10, 
                     search = "grid", 
                     returnData = TRUE,
                     returnResamp = "final",
                     savePredictions = "all",
                     classProbs = TRUE,
                     sampling = "smote",
                     summaryFunction = twoClassSummary,
                     selectionFunction = "best",
                     allowParallel = TRUE)

Answer 1

I reproduced the example of https://topepo.github.io/caret/subsampling-for-class-imbalances.html and tried another one. Results are:

perc.over = 200

perc.under = 200

It seems that the parameters are those of usage of the function SMOTE in the DMwR documentation (pag. 82):

Usage

SMOTE(form, data, perc.over = 200, k = 5, perc.under = 200,
    learner = NULL, ...)

In usage k = 5, but I verified only perc.over and perc.under:

library(caret)
library(DMwR)
        
set.seed(2969)
imbal_train <- twoClassSim(10000, intercept = -20, linearVars = 20)
table(imbal_train\\\$Class)
    
set.seed(9560)
smote_train <- SMOTE(Class ~ ., data  = imbal_train)                         
table(smote_train\\\$Class)[2]
        
perc.over <- 100*(table(smote_train\\\$Class)[2] - table(imbal_train\
    $Class)[2])/table(imbal_train\\\$Class)[2]
perc.under <- 100*table(smote_train\\\$Class)[1]/
    (table(smote_train\\\$Class)[2] - table(imbal_train\\\$Class)[2])
    
perc.over = 200

perc.under = 200
    
set.seed(1234)
imbal_train <- twoClassSim(10000, intercept = -40, linearVars = 40)

table(imbal_train\\\$Class)

set.seed(5678)
smote_train <- SMOTE(Class ~ ., data  = imbal_train)                         
table(smote_train\\\$Class)
        
perc.over <- 100*(table(smote_train\\\$Class)[2] - 
    table(imbal_train\\\$Class)[2])/table(imbal_train$Class)[2]
perc.under <- 100*table(smote_train\\\$Class)[1]/
    (table(smote_train\\\$Class)[2]-table(imbal_train\\\$Class)[2])
    
perc.over = 200

perc.under = 200

Answer 2

The package DMwR is being retired and should be replaced by the package themis.

My understanding is that we can't use SMOTE with weights in caret. So it's best to perform SMOTE separately and then use the balanced dataset to train in caret.